1 //===-- RuntimeDyldELF.cpp - Run-time dynamic linker for MC-JIT -*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // Implementation of ELF support for the MC-JIT runtime dynamic linker. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "RuntimeDyldELF.h" 14 #include "Targets/RuntimeDyldELFMips.h" 15 #include "llvm/ADT/StringRef.h" 16 #include "llvm/BinaryFormat/ELF.h" 17 #include "llvm/ExecutionEngine/Orc/SymbolStringPool.h" 18 #include "llvm/Object/ELFObjectFile.h" 19 #include "llvm/Object/ObjectFile.h" 20 #include "llvm/Support/Endian.h" 21 #include "llvm/Support/MemoryBuffer.h" 22 #include "llvm/TargetParser/Triple.h" 23 24 using namespace llvm; 25 using namespace llvm::object; 26 using namespace llvm::support::endian; 27 28 #define DEBUG_TYPE "dyld" 29 30 static void or32le(void *P, int32_t V) { write32le(P, read32le(P) | V); } 31 32 static void or32AArch64Imm(void *L, uint64_t Imm) { 33 or32le(L, (Imm & 0xFFF) << 10); 34 } 35 36 template <class T> static void write(bool isBE, void *P, T V) { 37 isBE ? write<T, llvm::endianness::big>(P, V) 38 : write<T, llvm::endianness::little>(P, V); 39 } 40 41 static void write32AArch64Addr(void *L, uint64_t Imm) { 42 uint32_t ImmLo = (Imm & 0x3) << 29; 43 uint32_t ImmHi = (Imm & 0x1FFFFC) << 3; 44 uint64_t Mask = (0x3 << 29) | (0x1FFFFC << 3); 45 write32le(L, (read32le(L) & ~Mask) | ImmLo | ImmHi); 46 } 47 48 // Return the bits [Start, End] from Val shifted Start bits. 49 // For instance, getBits(0xF0, 4, 8) returns 0xF. 50 static uint64_t getBits(uint64_t Val, int Start, int End) { 51 uint64_t Mask = ((uint64_t)1 << (End + 1 - Start)) - 1; 52 return (Val >> Start) & Mask; 53 } 54 55 namespace { 56 57 template <class ELFT> class DyldELFObject : public ELFObjectFile<ELFT> { 58 LLVM_ELF_IMPORT_TYPES_ELFT(ELFT) 59 60 typedef typename ELFT::uint addr_type; 61 62 DyldELFObject(ELFObjectFile<ELFT> &&Obj); 63 64 public: 65 static Expected<std::unique_ptr<DyldELFObject>> 66 create(MemoryBufferRef Wrapper); 67 68 void updateSectionAddress(const SectionRef &Sec, uint64_t Addr); 69 70 void updateSymbolAddress(const SymbolRef &SymRef, uint64_t Addr); 71 72 // Methods for type inquiry through isa, cast and dyn_cast 73 static bool classof(const Binary *v) { 74 return (isa<ELFObjectFile<ELFT>>(v) && 75 classof(cast<ELFObjectFile<ELFT>>(v))); 76 } 77 static bool classof(const ELFObjectFile<ELFT> *v) { 78 return v->isDyldType(); 79 } 80 }; 81 82 83 84 // The MemoryBuffer passed into this constructor is just a wrapper around the 85 // actual memory. Ultimately, the Binary parent class will take ownership of 86 // this MemoryBuffer object but not the underlying memory. 87 template <class ELFT> 88 DyldELFObject<ELFT>::DyldELFObject(ELFObjectFile<ELFT> &&Obj) 89 : ELFObjectFile<ELFT>(std::move(Obj)) { 90 this->isDyldELFObject = true; 91 } 92 93 template <class ELFT> 94 Expected<std::unique_ptr<DyldELFObject<ELFT>>> 95 DyldELFObject<ELFT>::create(MemoryBufferRef Wrapper) { 96 auto Obj = ELFObjectFile<ELFT>::create(Wrapper); 97 if (auto E = Obj.takeError()) 98 return std::move(E); 99 std::unique_ptr<DyldELFObject<ELFT>> Ret( 100 new DyldELFObject<ELFT>(std::move(*Obj))); 101 return std::move(Ret); 102 } 103 104 template <class ELFT> 105 void DyldELFObject<ELFT>::updateSectionAddress(const SectionRef &Sec, 106 uint64_t Addr) { 107 DataRefImpl ShdrRef = Sec.getRawDataRefImpl(); 108 Elf_Shdr *shdr = 109 const_cast<Elf_Shdr *>(reinterpret_cast<const Elf_Shdr *>(ShdrRef.p)); 110 111 // This assumes the address passed in matches the target address bitness 112 // The template-based type cast handles everything else. 113 shdr->sh_addr = static_cast<addr_type>(Addr); 114 } 115 116 template <class ELFT> 117 void DyldELFObject<ELFT>::updateSymbolAddress(const SymbolRef &SymRef, 118 uint64_t Addr) { 119 120 Elf_Sym *sym = const_cast<Elf_Sym *>( 121 ELFObjectFile<ELFT>::getSymbol(SymRef.getRawDataRefImpl())); 122 123 // This assumes the address passed in matches the target address bitness 124 // The template-based type cast handles everything else. 125 sym->st_value = static_cast<addr_type>(Addr); 126 } 127 128 class LoadedELFObjectInfo final 129 : public LoadedObjectInfoHelper<LoadedELFObjectInfo, 130 RuntimeDyld::LoadedObjectInfo> { 131 public: 132 LoadedELFObjectInfo(RuntimeDyldImpl &RTDyld, ObjSectionToIDMap ObjSecToIDMap) 133 : LoadedObjectInfoHelper(RTDyld, std::move(ObjSecToIDMap)) {} 134 135 OwningBinary<ObjectFile> 136 getObjectForDebug(const ObjectFile &Obj) const override; 137 }; 138 139 template <typename ELFT> 140 static Expected<std::unique_ptr<DyldELFObject<ELFT>>> 141 createRTDyldELFObject(MemoryBufferRef Buffer, const ObjectFile &SourceObject, 142 const LoadedELFObjectInfo &L) { 143 typedef typename ELFT::Shdr Elf_Shdr; 144 typedef typename ELFT::uint addr_type; 145 146 Expected<std::unique_ptr<DyldELFObject<ELFT>>> ObjOrErr = 147 DyldELFObject<ELFT>::create(Buffer); 148 if (Error E = ObjOrErr.takeError()) 149 return std::move(E); 150 151 std::unique_ptr<DyldELFObject<ELFT>> Obj = std::move(*ObjOrErr); 152 153 // Iterate over all sections in the object. 154 auto SI = SourceObject.section_begin(); 155 for (const auto &Sec : Obj->sections()) { 156 Expected<StringRef> NameOrErr = Sec.getName(); 157 if (!NameOrErr) { 158 consumeError(NameOrErr.takeError()); 159 continue; 160 } 161 162 if (*NameOrErr != "") { 163 DataRefImpl ShdrRef = Sec.getRawDataRefImpl(); 164 Elf_Shdr *shdr = const_cast<Elf_Shdr *>( 165 reinterpret_cast<const Elf_Shdr *>(ShdrRef.p)); 166 167 if (uint64_t SecLoadAddr = L.getSectionLoadAddress(*SI)) { 168 // This assumes that the address passed in matches the target address 169 // bitness. The template-based type cast handles everything else. 170 shdr->sh_addr = static_cast<addr_type>(SecLoadAddr); 171 } 172 } 173 ++SI; 174 } 175 176 return std::move(Obj); 177 } 178 179 static OwningBinary<ObjectFile> 180 createELFDebugObject(const ObjectFile &Obj, const LoadedELFObjectInfo &L) { 181 assert(Obj.isELF() && "Not an ELF object file."); 182 183 std::unique_ptr<MemoryBuffer> Buffer = 184 MemoryBuffer::getMemBufferCopy(Obj.getData(), Obj.getFileName()); 185 186 Expected<std::unique_ptr<ObjectFile>> DebugObj(nullptr); 187 handleAllErrors(DebugObj.takeError()); 188 if (Obj.getBytesInAddress() == 4 && Obj.isLittleEndian()) 189 DebugObj = 190 createRTDyldELFObject<ELF32LE>(Buffer->getMemBufferRef(), Obj, L); 191 else if (Obj.getBytesInAddress() == 4 && !Obj.isLittleEndian()) 192 DebugObj = 193 createRTDyldELFObject<ELF32BE>(Buffer->getMemBufferRef(), Obj, L); 194 else if (Obj.getBytesInAddress() == 8 && !Obj.isLittleEndian()) 195 DebugObj = 196 createRTDyldELFObject<ELF64BE>(Buffer->getMemBufferRef(), Obj, L); 197 else if (Obj.getBytesInAddress() == 8 && Obj.isLittleEndian()) 198 DebugObj = 199 createRTDyldELFObject<ELF64LE>(Buffer->getMemBufferRef(), Obj, L); 200 else 201 llvm_unreachable("Unexpected ELF format"); 202 203 handleAllErrors(DebugObj.takeError()); 204 return OwningBinary<ObjectFile>(std::move(*DebugObj), std::move(Buffer)); 205 } 206 207 OwningBinary<ObjectFile> 208 LoadedELFObjectInfo::getObjectForDebug(const ObjectFile &Obj) const { 209 return createELFDebugObject(Obj, *this); 210 } 211 212 } // anonymous namespace 213 214 namespace llvm { 215 216 RuntimeDyldELF::RuntimeDyldELF(RuntimeDyld::MemoryManager &MemMgr, 217 JITSymbolResolver &Resolver) 218 : RuntimeDyldImpl(MemMgr, Resolver), GOTSectionID(0), CurrentGOTIndex(0) {} 219 RuntimeDyldELF::~RuntimeDyldELF() = default; 220 221 void RuntimeDyldELF::registerEHFrames() { 222 for (SID EHFrameSID : UnregisteredEHFrameSections) { 223 uint8_t *EHFrameAddr = Sections[EHFrameSID].getAddress(); 224 uint64_t EHFrameLoadAddr = Sections[EHFrameSID].getLoadAddress(); 225 size_t EHFrameSize = Sections[EHFrameSID].getSize(); 226 MemMgr.registerEHFrames(EHFrameAddr, EHFrameLoadAddr, EHFrameSize); 227 } 228 UnregisteredEHFrameSections.clear(); 229 } 230 231 std::unique_ptr<RuntimeDyldELF> 232 llvm::RuntimeDyldELF::create(Triple::ArchType Arch, 233 RuntimeDyld::MemoryManager &MemMgr, 234 JITSymbolResolver &Resolver) { 235 switch (Arch) { 236 default: 237 return std::make_unique<RuntimeDyldELF>(MemMgr, Resolver); 238 case Triple::mips: 239 case Triple::mipsel: 240 case Triple::mips64: 241 case Triple::mips64el: 242 return std::make_unique<RuntimeDyldELFMips>(MemMgr, Resolver); 243 } 244 } 245 246 std::unique_ptr<RuntimeDyld::LoadedObjectInfo> 247 RuntimeDyldELF::loadObject(const object::ObjectFile &O) { 248 if (auto ObjSectionToIDOrErr = loadObjectImpl(O)) 249 return std::make_unique<LoadedELFObjectInfo>(*this, *ObjSectionToIDOrErr); 250 else { 251 HasError = true; 252 raw_string_ostream ErrStream(ErrorStr); 253 logAllUnhandledErrors(ObjSectionToIDOrErr.takeError(), ErrStream); 254 return nullptr; 255 } 256 } 257 258 void RuntimeDyldELF::resolveX86_64Relocation(const SectionEntry &Section, 259 uint64_t Offset, uint64_t Value, 260 uint32_t Type, int64_t Addend, 261 uint64_t SymOffset) { 262 switch (Type) { 263 default: 264 report_fatal_error("Relocation type not implemented yet!"); 265 break; 266 case ELF::R_X86_64_NONE: 267 break; 268 case ELF::R_X86_64_8: { 269 Value += Addend; 270 assert((int64_t)Value <= INT8_MAX && (int64_t)Value >= INT8_MIN); 271 uint8_t TruncatedAddr = (Value & 0xFF); 272 *Section.getAddressWithOffset(Offset) = TruncatedAddr; 273 LLVM_DEBUG(dbgs() << "Writing " << format("%p", TruncatedAddr) << " at " 274 << format("%p\n", Section.getAddressWithOffset(Offset))); 275 break; 276 } 277 case ELF::R_X86_64_16: { 278 Value += Addend; 279 assert((int64_t)Value <= INT16_MAX && (int64_t)Value >= INT16_MIN); 280 uint16_t TruncatedAddr = (Value & 0xFFFF); 281 support::ulittle16_t::ref(Section.getAddressWithOffset(Offset)) = 282 TruncatedAddr; 283 LLVM_DEBUG(dbgs() << "Writing " << format("%p", TruncatedAddr) << " at " 284 << format("%p\n", Section.getAddressWithOffset(Offset))); 285 break; 286 } 287 case ELF::R_X86_64_64: { 288 support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)) = 289 Value + Addend; 290 LLVM_DEBUG(dbgs() << "Writing " << format("%p", (Value + Addend)) << " at " 291 << format("%p\n", Section.getAddressWithOffset(Offset))); 292 break; 293 } 294 case ELF::R_X86_64_32: 295 case ELF::R_X86_64_32S: { 296 Value += Addend; 297 assert((Type == ELF::R_X86_64_32 && (Value <= UINT32_MAX)) || 298 (Type == ELF::R_X86_64_32S && 299 ((int64_t)Value <= INT32_MAX && (int64_t)Value >= INT32_MIN))); 300 uint32_t TruncatedAddr = (Value & 0xFFFFFFFF); 301 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 302 TruncatedAddr; 303 LLVM_DEBUG(dbgs() << "Writing " << format("%p", TruncatedAddr) << " at " 304 << format("%p\n", Section.getAddressWithOffset(Offset))); 305 break; 306 } 307 case ELF::R_X86_64_PC8: { 308 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 309 int64_t RealOffset = Value + Addend - FinalAddress; 310 assert(isInt<8>(RealOffset)); 311 int8_t TruncOffset = (RealOffset & 0xFF); 312 Section.getAddress()[Offset] = TruncOffset; 313 break; 314 } 315 case ELF::R_X86_64_PC32: { 316 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 317 int64_t RealOffset = Value + Addend - FinalAddress; 318 assert(isInt<32>(RealOffset)); 319 int32_t TruncOffset = (RealOffset & 0xFFFFFFFF); 320 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 321 TruncOffset; 322 break; 323 } 324 case ELF::R_X86_64_PC64: { 325 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 326 int64_t RealOffset = Value + Addend - FinalAddress; 327 support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)) = 328 RealOffset; 329 LLVM_DEBUG(dbgs() << "Writing " << format("%p", RealOffset) << " at " 330 << format("%p\n", FinalAddress)); 331 break; 332 } 333 case ELF::R_X86_64_GOTOFF64: { 334 // Compute Value - GOTBase. 335 uint64_t GOTBase = 0; 336 for (const auto &Section : Sections) { 337 if (Section.getName() == ".got") { 338 GOTBase = Section.getLoadAddressWithOffset(0); 339 break; 340 } 341 } 342 assert(GOTBase != 0 && "missing GOT"); 343 int64_t GOTOffset = Value - GOTBase + Addend; 344 support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)) = GOTOffset; 345 break; 346 } 347 case ELF::R_X86_64_DTPMOD64: { 348 // We only have one DSO, so the module id is always 1. 349 support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)) = 1; 350 break; 351 } 352 case ELF::R_X86_64_DTPOFF64: 353 case ELF::R_X86_64_TPOFF64: { 354 // DTPOFF64 should resolve to the offset in the TLS block, TPOFF64 to the 355 // offset in the *initial* TLS block. Since we are statically linking, all 356 // TLS blocks already exist in the initial block, so resolve both 357 // relocations equally. 358 support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)) = 359 Value + Addend; 360 break; 361 } 362 case ELF::R_X86_64_DTPOFF32: 363 case ELF::R_X86_64_TPOFF32: { 364 // As for the (D)TPOFF64 relocations above, both DTPOFF32 and TPOFF32 can 365 // be resolved equally. 366 int64_t RealValue = Value + Addend; 367 assert(RealValue >= INT32_MIN && RealValue <= INT32_MAX); 368 int32_t TruncValue = RealValue; 369 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 370 TruncValue; 371 break; 372 } 373 } 374 } 375 376 void RuntimeDyldELF::resolveX86Relocation(const SectionEntry &Section, 377 uint64_t Offset, uint32_t Value, 378 uint32_t Type, int32_t Addend) { 379 switch (Type) { 380 case ELF::R_386_32: { 381 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 382 Value + Addend; 383 break; 384 } 385 // Handle R_386_PLT32 like R_386_PC32 since it should be able to 386 // reach any 32 bit address. 387 case ELF::R_386_PLT32: 388 case ELF::R_386_PC32: { 389 uint32_t FinalAddress = 390 Section.getLoadAddressWithOffset(Offset) & 0xFFFFFFFF; 391 uint32_t RealOffset = Value + Addend - FinalAddress; 392 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 393 RealOffset; 394 break; 395 } 396 default: 397 // There are other relocation types, but it appears these are the 398 // only ones currently used by the LLVM ELF object writer 399 report_fatal_error("Relocation type not implemented yet!"); 400 break; 401 } 402 } 403 404 void RuntimeDyldELF::resolveAArch64Relocation(const SectionEntry &Section, 405 uint64_t Offset, uint64_t Value, 406 uint32_t Type, int64_t Addend) { 407 uint32_t *TargetPtr = 408 reinterpret_cast<uint32_t *>(Section.getAddressWithOffset(Offset)); 409 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 410 // Data should use target endian. Code should always use little endian. 411 bool isBE = Arch == Triple::aarch64_be; 412 413 LLVM_DEBUG(dbgs() << "resolveAArch64Relocation, LocalAddress: 0x" 414 << format("%llx", Section.getAddressWithOffset(Offset)) 415 << " FinalAddress: 0x" << format("%llx", FinalAddress) 416 << " Value: 0x" << format("%llx", Value) << " Type: 0x" 417 << format("%x", Type) << " Addend: 0x" 418 << format("%llx", Addend) << "\n"); 419 420 switch (Type) { 421 default: 422 report_fatal_error("Relocation type not implemented yet!"); 423 break; 424 case ELF::R_AARCH64_NONE: 425 break; 426 case ELF::R_AARCH64_ABS16: { 427 uint64_t Result = Value + Addend; 428 assert(Result == static_cast<uint64_t>(llvm::SignExtend64(Result, 16)) || 429 (Result >> 16) == 0); 430 write(isBE, TargetPtr, static_cast<uint16_t>(Result & 0xffffU)); 431 break; 432 } 433 case ELF::R_AARCH64_ABS32: { 434 uint64_t Result = Value + Addend; 435 assert(Result == static_cast<uint64_t>(llvm::SignExtend64(Result, 32)) || 436 (Result >> 32) == 0); 437 write(isBE, TargetPtr, static_cast<uint32_t>(Result & 0xffffffffU)); 438 break; 439 } 440 case ELF::R_AARCH64_ABS64: 441 write(isBE, TargetPtr, Value + Addend); 442 break; 443 case ELF::R_AARCH64_PLT32: { 444 uint64_t Result = Value + Addend - FinalAddress; 445 assert(static_cast<int64_t>(Result) >= INT32_MIN && 446 static_cast<int64_t>(Result) <= INT32_MAX); 447 write(isBE, TargetPtr, static_cast<uint32_t>(Result)); 448 break; 449 } 450 case ELF::R_AARCH64_PREL16: { 451 uint64_t Result = Value + Addend - FinalAddress; 452 assert(static_cast<int64_t>(Result) >= INT16_MIN && 453 static_cast<int64_t>(Result) <= UINT16_MAX); 454 write(isBE, TargetPtr, static_cast<uint16_t>(Result & 0xffffU)); 455 break; 456 } 457 case ELF::R_AARCH64_PREL32: { 458 uint64_t Result = Value + Addend - FinalAddress; 459 assert(static_cast<int64_t>(Result) >= INT32_MIN && 460 static_cast<int64_t>(Result) <= UINT32_MAX); 461 write(isBE, TargetPtr, static_cast<uint32_t>(Result & 0xffffffffU)); 462 break; 463 } 464 case ELF::R_AARCH64_PREL64: 465 write(isBE, TargetPtr, Value + Addend - FinalAddress); 466 break; 467 case ELF::R_AARCH64_CONDBR19: { 468 uint64_t BranchImm = Value + Addend - FinalAddress; 469 470 assert(isInt<21>(BranchImm)); 471 *TargetPtr &= 0xff00001fU; 472 // Immediate:20:2 goes in bits 23:5 of Bcc, CBZ, CBNZ 473 or32le(TargetPtr, (BranchImm & 0x001FFFFC) << 3); 474 break; 475 } 476 case ELF::R_AARCH64_TSTBR14: { 477 uint64_t BranchImm = Value + Addend - FinalAddress; 478 479 assert(isInt<16>(BranchImm)); 480 481 uint32_t RawInstr = *(support::little32_t *)TargetPtr; 482 *(support::little32_t *)TargetPtr = RawInstr & 0xfff8001fU; 483 484 // Immediate:15:2 goes in bits 18:5 of TBZ, TBNZ 485 or32le(TargetPtr, (BranchImm & 0x0000FFFC) << 3); 486 break; 487 } 488 case ELF::R_AARCH64_CALL26: // fallthrough 489 case ELF::R_AARCH64_JUMP26: { 490 // Operation: S+A-P. Set Call or B immediate value to bits fff_fffc of the 491 // calculation. 492 uint64_t BranchImm = Value + Addend - FinalAddress; 493 494 // "Check that -2^27 <= result < 2^27". 495 assert(isInt<28>(BranchImm)); 496 or32le(TargetPtr, (BranchImm & 0x0FFFFFFC) >> 2); 497 break; 498 } 499 case ELF::R_AARCH64_MOVW_UABS_G3: 500 or32le(TargetPtr, ((Value + Addend) & 0xFFFF000000000000) >> 43); 501 break; 502 case ELF::R_AARCH64_MOVW_UABS_G2_NC: 503 or32le(TargetPtr, ((Value + Addend) & 0xFFFF00000000) >> 27); 504 break; 505 case ELF::R_AARCH64_MOVW_UABS_G1_NC: 506 or32le(TargetPtr, ((Value + Addend) & 0xFFFF0000) >> 11); 507 break; 508 case ELF::R_AARCH64_MOVW_UABS_G0_NC: 509 or32le(TargetPtr, ((Value + Addend) & 0xFFFF) << 5); 510 break; 511 case ELF::R_AARCH64_ADR_PREL_PG_HI21: { 512 // Operation: Page(S+A) - Page(P) 513 uint64_t Result = 514 ((Value + Addend) & ~0xfffULL) - (FinalAddress & ~0xfffULL); 515 516 // Check that -2^32 <= X < 2^32 517 assert(isInt<33>(Result) && "overflow check failed for relocation"); 518 519 // Immediate goes in bits 30:29 + 5:23 of ADRP instruction, taken 520 // from bits 32:12 of X. 521 write32AArch64Addr(TargetPtr, Result >> 12); 522 break; 523 } 524 case ELF::R_AARCH64_ADD_ABS_LO12_NC: 525 // Operation: S + A 526 // Immediate goes in bits 21:10 of LD/ST instruction, taken 527 // from bits 11:0 of X 528 or32AArch64Imm(TargetPtr, Value + Addend); 529 break; 530 case ELF::R_AARCH64_LDST8_ABS_LO12_NC: 531 // Operation: S + A 532 // Immediate goes in bits 21:10 of LD/ST instruction, taken 533 // from bits 11:0 of X 534 or32AArch64Imm(TargetPtr, getBits(Value + Addend, 0, 11)); 535 break; 536 case ELF::R_AARCH64_LDST16_ABS_LO12_NC: 537 // Operation: S + A 538 // Immediate goes in bits 21:10 of LD/ST instruction, taken 539 // from bits 11:1 of X 540 or32AArch64Imm(TargetPtr, getBits(Value + Addend, 1, 11)); 541 break; 542 case ELF::R_AARCH64_LDST32_ABS_LO12_NC: 543 // Operation: S + A 544 // Immediate goes in bits 21:10 of LD/ST instruction, taken 545 // from bits 11:2 of X 546 or32AArch64Imm(TargetPtr, getBits(Value + Addend, 2, 11)); 547 break; 548 case ELF::R_AARCH64_LDST64_ABS_LO12_NC: 549 // Operation: S + A 550 // Immediate goes in bits 21:10 of LD/ST instruction, taken 551 // from bits 11:3 of X 552 or32AArch64Imm(TargetPtr, getBits(Value + Addend, 3, 11)); 553 break; 554 case ELF::R_AARCH64_LDST128_ABS_LO12_NC: 555 // Operation: S + A 556 // Immediate goes in bits 21:10 of LD/ST instruction, taken 557 // from bits 11:4 of X 558 or32AArch64Imm(TargetPtr, getBits(Value + Addend, 4, 11)); 559 break; 560 case ELF::R_AARCH64_LD_PREL_LO19: { 561 // Operation: S + A - P 562 uint64_t Result = Value + Addend - FinalAddress; 563 564 // "Check that -2^20 <= result < 2^20". 565 assert(isInt<21>(Result)); 566 567 *TargetPtr &= 0xff00001fU; 568 // Immediate goes in bits 23:5 of LD imm instruction, taken 569 // from bits 20:2 of X 570 *TargetPtr |= ((Result & 0xffc) << (5 - 2)); 571 break; 572 } 573 case ELF::R_AARCH64_ADR_PREL_LO21: { 574 // Operation: S + A - P 575 uint64_t Result = Value + Addend - FinalAddress; 576 577 // "Check that -2^20 <= result < 2^20". 578 assert(isInt<21>(Result)); 579 580 *TargetPtr &= 0x9f00001fU; 581 // Immediate goes in bits 23:5, 30:29 of ADR imm instruction, taken 582 // from bits 20:0 of X 583 *TargetPtr |= ((Result & 0xffc) << (5 - 2)); 584 *TargetPtr |= (Result & 0x3) << 29; 585 break; 586 } 587 } 588 } 589 590 void RuntimeDyldELF::resolveARMRelocation(const SectionEntry &Section, 591 uint64_t Offset, uint32_t Value, 592 uint32_t Type, int32_t Addend) { 593 // TODO: Add Thumb relocations. 594 uint32_t *TargetPtr = 595 reinterpret_cast<uint32_t *>(Section.getAddressWithOffset(Offset)); 596 uint32_t FinalAddress = Section.getLoadAddressWithOffset(Offset) & 0xFFFFFFFF; 597 Value += Addend; 598 599 LLVM_DEBUG(dbgs() << "resolveARMRelocation, LocalAddress: " 600 << Section.getAddressWithOffset(Offset) 601 << " FinalAddress: " << format("%p", FinalAddress) 602 << " Value: " << format("%x", Value) 603 << " Type: " << format("%x", Type) 604 << " Addend: " << format("%x", Addend) << "\n"); 605 606 switch (Type) { 607 default: 608 llvm_unreachable("Not implemented relocation type!"); 609 610 case ELF::R_ARM_NONE: 611 break; 612 // Write a 31bit signed offset 613 case ELF::R_ARM_PREL31: 614 support::ulittle32_t::ref{TargetPtr} = 615 (support::ulittle32_t::ref{TargetPtr} & 0x80000000) | 616 ((Value - FinalAddress) & ~0x80000000); 617 break; 618 case ELF::R_ARM_TARGET1: 619 case ELF::R_ARM_ABS32: 620 support::ulittle32_t::ref{TargetPtr} = Value; 621 break; 622 // Write first 16 bit of 32 bit value to the mov instruction. 623 // Last 4 bit should be shifted. 624 case ELF::R_ARM_MOVW_ABS_NC: 625 case ELF::R_ARM_MOVT_ABS: 626 if (Type == ELF::R_ARM_MOVW_ABS_NC) 627 Value = Value & 0xFFFF; 628 else if (Type == ELF::R_ARM_MOVT_ABS) 629 Value = (Value >> 16) & 0xFFFF; 630 support::ulittle32_t::ref{TargetPtr} = 631 (support::ulittle32_t::ref{TargetPtr} & ~0x000F0FFF) | (Value & 0xFFF) | 632 (((Value >> 12) & 0xF) << 16); 633 break; 634 // Write 24 bit relative value to the branch instruction. 635 case ELF::R_ARM_PC24: // Fall through. 636 case ELF::R_ARM_CALL: // Fall through. 637 case ELF::R_ARM_JUMP24: 638 int32_t RelValue = static_cast<int32_t>(Value - FinalAddress - 8); 639 RelValue = (RelValue & 0x03FFFFFC) >> 2; 640 assert((support::ulittle32_t::ref{TargetPtr} & 0xFFFFFF) == 0xFFFFFE); 641 support::ulittle32_t::ref{TargetPtr} = 642 (support::ulittle32_t::ref{TargetPtr} & 0xFF000000) | RelValue; 643 break; 644 } 645 } 646 647 bool RuntimeDyldELF::resolveLoongArch64ShortBranch( 648 unsigned SectionID, relocation_iterator RelI, 649 const RelocationValueRef &Value) { 650 uint64_t Address; 651 if (Value.SymbolName) { 652 auto Loc = GlobalSymbolTable.find(Value.SymbolName); 653 // Don't create direct branch for external symbols. 654 if (Loc == GlobalSymbolTable.end()) 655 return false; 656 const auto &SymInfo = Loc->second; 657 Address = 658 uint64_t(Sections[SymInfo.getSectionID()].getLoadAddressWithOffset( 659 SymInfo.getOffset())); 660 } else { 661 Address = uint64_t(Sections[Value.SectionID].getLoadAddress()); 662 } 663 uint64_t Offset = RelI->getOffset(); 664 uint64_t SourceAddress = Sections[SectionID].getLoadAddressWithOffset(Offset); 665 if (!isInt<28>(Address + Value.Addend - SourceAddress)) 666 return false; 667 resolveRelocation(Sections[SectionID], Offset, Address, RelI->getType(), 668 Value.Addend); 669 return true; 670 } 671 672 void RuntimeDyldELF::resolveLoongArch64Branch(unsigned SectionID, 673 const RelocationValueRef &Value, 674 relocation_iterator RelI, 675 StubMap &Stubs) { 676 LLVM_DEBUG(dbgs() << "\t\tThis is an LoongArch64 branch relocation.\n"); 677 678 if (resolveLoongArch64ShortBranch(SectionID, RelI, Value)) 679 return; 680 681 SectionEntry &Section = Sections[SectionID]; 682 uint64_t Offset = RelI->getOffset(); 683 unsigned RelType = RelI->getType(); 684 // Look for an existing stub. 685 StubMap::const_iterator i = Stubs.find(Value); 686 if (i != Stubs.end()) { 687 resolveRelocation(Section, Offset, 688 (uint64_t)Section.getAddressWithOffset(i->second), 689 RelType, 0); 690 LLVM_DEBUG(dbgs() << " Stub function found\n"); 691 return; 692 } 693 // Create a new stub function. 694 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 695 Stubs[Value] = Section.getStubOffset(); 696 uint8_t *StubTargetAddr = 697 createStubFunction(Section.getAddressWithOffset(Section.getStubOffset())); 698 RelocationEntry LU12I_W(SectionID, StubTargetAddr - Section.getAddress(), 699 ELF::R_LARCH_ABS_HI20, Value.Addend); 700 RelocationEntry ORI(SectionID, StubTargetAddr - Section.getAddress() + 4, 701 ELF::R_LARCH_ABS_LO12, Value.Addend); 702 RelocationEntry LU32I_D(SectionID, StubTargetAddr - Section.getAddress() + 8, 703 ELF::R_LARCH_ABS64_LO20, Value.Addend); 704 RelocationEntry LU52I_D(SectionID, StubTargetAddr - Section.getAddress() + 12, 705 ELF::R_LARCH_ABS64_HI12, Value.Addend); 706 if (Value.SymbolName) { 707 addRelocationForSymbol(LU12I_W, Value.SymbolName); 708 addRelocationForSymbol(ORI, Value.SymbolName); 709 addRelocationForSymbol(LU32I_D, Value.SymbolName); 710 addRelocationForSymbol(LU52I_D, Value.SymbolName); 711 } else { 712 addRelocationForSection(LU12I_W, Value.SectionID); 713 addRelocationForSection(ORI, Value.SectionID); 714 addRelocationForSection(LU32I_D, Value.SectionID); 715 716 addRelocationForSection(LU52I_D, Value.SectionID); 717 } 718 resolveRelocation(Section, Offset, 719 reinterpret_cast<uint64_t>( 720 Section.getAddressWithOffset(Section.getStubOffset())), 721 RelType, 0); 722 Section.advanceStubOffset(getMaxStubSize()); 723 } 724 725 // Returns extract bits Val[Hi:Lo]. 726 static inline uint32_t extractBits(uint64_t Val, uint32_t Hi, uint32_t Lo) { 727 return Hi == 63 ? Val >> Lo : (Val & (((1ULL << (Hi + 1)) - 1))) >> Lo; 728 } 729 730 void RuntimeDyldELF::resolveLoongArch64Relocation(const SectionEntry &Section, 731 uint64_t Offset, 732 uint64_t Value, uint32_t Type, 733 int64_t Addend) { 734 auto *TargetPtr = Section.getAddressWithOffset(Offset); 735 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 736 737 LLVM_DEBUG(dbgs() << "resolveLoongArch64Relocation, LocalAddress: 0x" 738 << format("%llx", Section.getAddressWithOffset(Offset)) 739 << " FinalAddress: 0x" << format("%llx", FinalAddress) 740 << " Value: 0x" << format("%llx", Value) << " Type: 0x" 741 << format("%x", Type) << " Addend: 0x" 742 << format("%llx", Addend) << "\n"); 743 744 switch (Type) { 745 default: 746 report_fatal_error("Relocation type not implemented yet!"); 747 break; 748 case ELF::R_LARCH_32: 749 support::ulittle32_t::ref{TargetPtr} = 750 static_cast<uint32_t>(Value + Addend); 751 break; 752 case ELF::R_LARCH_64: 753 support::ulittle64_t::ref{TargetPtr} = Value + Addend; 754 break; 755 case ELF::R_LARCH_32_PCREL: 756 support::ulittle32_t::ref{TargetPtr} = 757 static_cast<uint32_t>(Value + Addend - FinalAddress); 758 break; 759 case ELF::R_LARCH_B26: { 760 uint64_t B26 = (Value + Addend - FinalAddress) >> 2; 761 auto Instr = support::ulittle32_t::ref(TargetPtr); 762 uint32_t Imm15_0 = extractBits(B26, /*Hi=*/15, /*Lo=*/0) << 10; 763 uint32_t Imm25_16 = extractBits(B26, /*Hi=*/25, /*Lo=*/16); 764 Instr = (Instr & 0xfc000000) | Imm15_0 | Imm25_16; 765 break; 766 } 767 case ELF::R_LARCH_CALL36: { 768 uint64_t Call36 = (Value + Addend - FinalAddress) >> 2; 769 auto Pcaddu18i = support::ulittle32_t::ref(TargetPtr); 770 uint32_t Imm35_16 = 771 extractBits((Call36 + (1UL << 15)), /*Hi=*/35, /*Lo=*/16) << 5; 772 Pcaddu18i = (Pcaddu18i & 0xfe00001f) | Imm35_16; 773 auto Jirl = support::ulittle32_t::ref(TargetPtr + 4); 774 uint32_t Imm15_0 = extractBits(Call36, /*Hi=*/15, /*Lo=*/0) << 10; 775 Jirl = (Jirl & 0xfc0003ff) | Imm15_0; 776 break; 777 } 778 case ELF::R_LARCH_GOT_PC_HI20: 779 case ELF::R_LARCH_PCALA_HI20: { 780 uint64_t Target = Value + Addend; 781 uint64_t TargetPage = 782 (Target + (Target & 0x800)) & ~static_cast<uint64_t>(0xfff); 783 uint64_t PCPage = FinalAddress & ~static_cast<uint64_t>(0xfff); 784 int64_t PageDelta = TargetPage - PCPage; 785 auto Instr = support::ulittle32_t::ref(TargetPtr); 786 uint32_t Imm31_12 = extractBits(PageDelta, /*Hi=*/31, /*Lo=*/12) << 5; 787 Instr = (Instr & 0xfe00001f) | Imm31_12; 788 break; 789 } 790 case ELF::R_LARCH_GOT_PC_LO12: 791 case ELF::R_LARCH_PCALA_LO12: { 792 uint64_t TargetOffset = (Value + Addend) & 0xfff; 793 auto Instr = support::ulittle32_t::ref(TargetPtr); 794 uint32_t Imm11_0 = TargetOffset << 10; 795 Instr = (Instr & 0xffc003ff) | Imm11_0; 796 break; 797 } 798 case ELF::R_LARCH_ABS_HI20: { 799 uint64_t Target = Value + Addend; 800 auto Instr = support::ulittle32_t::ref(TargetPtr); 801 uint32_t Imm31_12 = extractBits(Target, /*Hi=*/31, /*Lo=*/12) << 5; 802 Instr = (Instr & 0xfe00001f) | Imm31_12; 803 break; 804 } 805 case ELF::R_LARCH_ABS_LO12: { 806 uint64_t Target = Value + Addend; 807 auto Instr = support::ulittle32_t::ref(TargetPtr); 808 uint32_t Imm11_0 = extractBits(Target, /*Hi=*/11, /*Lo=*/0) << 10; 809 Instr = (Instr & 0xffc003ff) | Imm11_0; 810 break; 811 } 812 case ELF::R_LARCH_ABS64_LO20: { 813 uint64_t Target = Value + Addend; 814 auto Instr = support::ulittle32_t::ref(TargetPtr); 815 uint32_t Imm51_32 = extractBits(Target, /*Hi=*/51, /*Lo=*/32) << 5; 816 Instr = (Instr & 0xfe00001f) | Imm51_32; 817 break; 818 } 819 case ELF::R_LARCH_ABS64_HI12: { 820 uint64_t Target = Value + Addend; 821 auto Instr = support::ulittle32_t::ref(TargetPtr); 822 uint32_t Imm63_52 = extractBits(Target, /*Hi=*/63, /*Lo=*/52) << 10; 823 Instr = (Instr & 0xffc003ff) | Imm63_52; 824 break; 825 } 826 case ELF::R_LARCH_ADD32: 827 support::ulittle32_t::ref{TargetPtr} = 828 (support::ulittle32_t::ref{TargetPtr} + 829 static_cast<uint32_t>(Value + Addend)); 830 break; 831 case ELF::R_LARCH_SUB32: 832 support::ulittle32_t::ref{TargetPtr} = 833 (support::ulittle32_t::ref{TargetPtr} - 834 static_cast<uint32_t>(Value + Addend)); 835 break; 836 case ELF::R_LARCH_ADD64: 837 support::ulittle64_t::ref{TargetPtr} = 838 (support::ulittle64_t::ref{TargetPtr} + Value + Addend); 839 break; 840 case ELF::R_LARCH_SUB64: 841 support::ulittle64_t::ref{TargetPtr} = 842 (support::ulittle64_t::ref{TargetPtr} - Value - Addend); 843 break; 844 } 845 } 846 847 void RuntimeDyldELF::setMipsABI(const ObjectFile &Obj) { 848 if (Arch == Triple::UnknownArch || 849 Triple::getArchTypePrefix(Arch) != "mips") { 850 IsMipsO32ABI = false; 851 IsMipsN32ABI = false; 852 IsMipsN64ABI = false; 853 return; 854 } 855 if (auto *E = dyn_cast<ELFObjectFileBase>(&Obj)) { 856 unsigned AbiVariant = E->getPlatformFlags(); 857 IsMipsO32ABI = AbiVariant & ELF::EF_MIPS_ABI_O32; 858 IsMipsN32ABI = AbiVariant & ELF::EF_MIPS_ABI2; 859 } 860 IsMipsN64ABI = Obj.getFileFormatName() == "elf64-mips"; 861 } 862 863 // Return the .TOC. section and offset. 864 Error RuntimeDyldELF::findPPC64TOCSection(const ELFObjectFileBase &Obj, 865 ObjSectionToIDMap &LocalSections, 866 RelocationValueRef &Rel) { 867 // Set a default SectionID in case we do not find a TOC section below. 868 // This may happen for references to TOC base base (sym@toc, .odp 869 // relocation) without a .toc directive. In this case just use the 870 // first section (which is usually the .odp) since the code won't 871 // reference the .toc base directly. 872 Rel.SymbolName = nullptr; 873 Rel.SectionID = 0; 874 875 // The TOC consists of sections .got, .toc, .tocbss, .plt in that 876 // order. The TOC starts where the first of these sections starts. 877 for (auto &Section : Obj.sections()) { 878 Expected<StringRef> NameOrErr = Section.getName(); 879 if (!NameOrErr) 880 return NameOrErr.takeError(); 881 StringRef SectionName = *NameOrErr; 882 883 if (SectionName == ".got" 884 || SectionName == ".toc" 885 || SectionName == ".tocbss" 886 || SectionName == ".plt") { 887 if (auto SectionIDOrErr = 888 findOrEmitSection(Obj, Section, false, LocalSections)) 889 Rel.SectionID = *SectionIDOrErr; 890 else 891 return SectionIDOrErr.takeError(); 892 break; 893 } 894 } 895 896 // Per the ppc64-elf-linux ABI, The TOC base is TOC value plus 0x8000 897 // thus permitting a full 64 Kbytes segment. 898 Rel.Addend = 0x8000; 899 900 return Error::success(); 901 } 902 903 // Returns the sections and offset associated with the ODP entry referenced 904 // by Symbol. 905 Error RuntimeDyldELF::findOPDEntrySection(const ELFObjectFileBase &Obj, 906 ObjSectionToIDMap &LocalSections, 907 RelocationValueRef &Rel) { 908 // Get the ELF symbol value (st_value) to compare with Relocation offset in 909 // .opd entries 910 for (section_iterator si = Obj.section_begin(), se = Obj.section_end(); 911 si != se; ++si) { 912 913 Expected<section_iterator> RelSecOrErr = si->getRelocatedSection(); 914 if (!RelSecOrErr) 915 report_fatal_error(Twine(toString(RelSecOrErr.takeError()))); 916 917 section_iterator RelSecI = *RelSecOrErr; 918 if (RelSecI == Obj.section_end()) 919 continue; 920 921 Expected<StringRef> NameOrErr = RelSecI->getName(); 922 if (!NameOrErr) 923 return NameOrErr.takeError(); 924 StringRef RelSectionName = *NameOrErr; 925 926 if (RelSectionName != ".opd") 927 continue; 928 929 for (elf_relocation_iterator i = si->relocation_begin(), 930 e = si->relocation_end(); 931 i != e;) { 932 // The R_PPC64_ADDR64 relocation indicates the first field 933 // of a .opd entry 934 uint64_t TypeFunc = i->getType(); 935 if (TypeFunc != ELF::R_PPC64_ADDR64) { 936 ++i; 937 continue; 938 } 939 940 uint64_t TargetSymbolOffset = i->getOffset(); 941 symbol_iterator TargetSymbol = i->getSymbol(); 942 int64_t Addend; 943 if (auto AddendOrErr = i->getAddend()) 944 Addend = *AddendOrErr; 945 else 946 return AddendOrErr.takeError(); 947 948 ++i; 949 if (i == e) 950 break; 951 952 // Just check if following relocation is a R_PPC64_TOC 953 uint64_t TypeTOC = i->getType(); 954 if (TypeTOC != ELF::R_PPC64_TOC) 955 continue; 956 957 // Finally compares the Symbol value and the target symbol offset 958 // to check if this .opd entry refers to the symbol the relocation 959 // points to. 960 if (Rel.Addend != (int64_t)TargetSymbolOffset) 961 continue; 962 963 section_iterator TSI = Obj.section_end(); 964 if (auto TSIOrErr = TargetSymbol->getSection()) 965 TSI = *TSIOrErr; 966 else 967 return TSIOrErr.takeError(); 968 assert(TSI != Obj.section_end() && "TSI should refer to a valid section"); 969 970 bool IsCode = TSI->isText(); 971 if (auto SectionIDOrErr = findOrEmitSection(Obj, *TSI, IsCode, 972 LocalSections)) 973 Rel.SectionID = *SectionIDOrErr; 974 else 975 return SectionIDOrErr.takeError(); 976 Rel.Addend = (intptr_t)Addend; 977 return Error::success(); 978 } 979 } 980 llvm_unreachable("Attempting to get address of ODP entry!"); 981 } 982 983 // Relocation masks following the #lo(value), #hi(value), #ha(value), 984 // #higher(value), #highera(value), #highest(value), and #highesta(value) 985 // macros defined in section 4.5.1. Relocation Types of the PPC-elf64abi 986 // document. 987 988 static inline uint16_t applyPPClo(uint64_t value) { return value & 0xffff; } 989 990 static inline uint16_t applyPPChi(uint64_t value) { 991 return (value >> 16) & 0xffff; 992 } 993 994 static inline uint16_t applyPPCha (uint64_t value) { 995 return ((value + 0x8000) >> 16) & 0xffff; 996 } 997 998 static inline uint16_t applyPPChigher(uint64_t value) { 999 return (value >> 32) & 0xffff; 1000 } 1001 1002 static inline uint16_t applyPPChighera (uint64_t value) { 1003 return ((value + 0x8000) >> 32) & 0xffff; 1004 } 1005 1006 static inline uint16_t applyPPChighest(uint64_t value) { 1007 return (value >> 48) & 0xffff; 1008 } 1009 1010 static inline uint16_t applyPPChighesta (uint64_t value) { 1011 return ((value + 0x8000) >> 48) & 0xffff; 1012 } 1013 1014 void RuntimeDyldELF::resolvePPC32Relocation(const SectionEntry &Section, 1015 uint64_t Offset, uint64_t Value, 1016 uint32_t Type, int64_t Addend) { 1017 uint8_t *LocalAddress = Section.getAddressWithOffset(Offset); 1018 switch (Type) { 1019 default: 1020 report_fatal_error("Relocation type not implemented yet!"); 1021 break; 1022 case ELF::R_PPC_ADDR16_LO: 1023 writeInt16BE(LocalAddress, applyPPClo(Value + Addend)); 1024 break; 1025 case ELF::R_PPC_ADDR16_HI: 1026 writeInt16BE(LocalAddress, applyPPChi(Value + Addend)); 1027 break; 1028 case ELF::R_PPC_ADDR16_HA: 1029 writeInt16BE(LocalAddress, applyPPCha(Value + Addend)); 1030 break; 1031 } 1032 } 1033 1034 void RuntimeDyldELF::resolvePPC64Relocation(const SectionEntry &Section, 1035 uint64_t Offset, uint64_t Value, 1036 uint32_t Type, int64_t Addend) { 1037 uint8_t *LocalAddress = Section.getAddressWithOffset(Offset); 1038 switch (Type) { 1039 default: 1040 report_fatal_error("Relocation type not implemented yet!"); 1041 break; 1042 case ELF::R_PPC64_ADDR16: 1043 writeInt16BE(LocalAddress, applyPPClo(Value + Addend)); 1044 break; 1045 case ELF::R_PPC64_ADDR16_DS: 1046 writeInt16BE(LocalAddress, applyPPClo(Value + Addend) & ~3); 1047 break; 1048 case ELF::R_PPC64_ADDR16_LO: 1049 writeInt16BE(LocalAddress, applyPPClo(Value + Addend)); 1050 break; 1051 case ELF::R_PPC64_ADDR16_LO_DS: 1052 writeInt16BE(LocalAddress, applyPPClo(Value + Addend) & ~3); 1053 break; 1054 case ELF::R_PPC64_ADDR16_HI: 1055 case ELF::R_PPC64_ADDR16_HIGH: 1056 writeInt16BE(LocalAddress, applyPPChi(Value + Addend)); 1057 break; 1058 case ELF::R_PPC64_ADDR16_HA: 1059 case ELF::R_PPC64_ADDR16_HIGHA: 1060 writeInt16BE(LocalAddress, applyPPCha(Value + Addend)); 1061 break; 1062 case ELF::R_PPC64_ADDR16_HIGHER: 1063 writeInt16BE(LocalAddress, applyPPChigher(Value + Addend)); 1064 break; 1065 case ELF::R_PPC64_ADDR16_HIGHERA: 1066 writeInt16BE(LocalAddress, applyPPChighera(Value + Addend)); 1067 break; 1068 case ELF::R_PPC64_ADDR16_HIGHEST: 1069 writeInt16BE(LocalAddress, applyPPChighest(Value + Addend)); 1070 break; 1071 case ELF::R_PPC64_ADDR16_HIGHESTA: 1072 writeInt16BE(LocalAddress, applyPPChighesta(Value + Addend)); 1073 break; 1074 case ELF::R_PPC64_ADDR14: { 1075 assert(((Value + Addend) & 3) == 0); 1076 // Preserve the AA/LK bits in the branch instruction 1077 uint8_t aalk = *(LocalAddress + 3); 1078 writeInt16BE(LocalAddress + 2, (aalk & 3) | ((Value + Addend) & 0xfffc)); 1079 } break; 1080 case ELF::R_PPC64_REL16_LO: { 1081 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1082 uint64_t Delta = Value - FinalAddress + Addend; 1083 writeInt16BE(LocalAddress, applyPPClo(Delta)); 1084 } break; 1085 case ELF::R_PPC64_REL16_HI: { 1086 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1087 uint64_t Delta = Value - FinalAddress + Addend; 1088 writeInt16BE(LocalAddress, applyPPChi(Delta)); 1089 } break; 1090 case ELF::R_PPC64_REL16_HA: { 1091 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1092 uint64_t Delta = Value - FinalAddress + Addend; 1093 writeInt16BE(LocalAddress, applyPPCha(Delta)); 1094 } break; 1095 case ELF::R_PPC64_ADDR32: { 1096 int64_t Result = static_cast<int64_t>(Value + Addend); 1097 if (SignExtend64<32>(Result) != Result) 1098 llvm_unreachable("Relocation R_PPC64_ADDR32 overflow"); 1099 writeInt32BE(LocalAddress, Result); 1100 } break; 1101 case ELF::R_PPC64_REL24: { 1102 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1103 int64_t delta = static_cast<int64_t>(Value - FinalAddress + Addend); 1104 if (SignExtend64<26>(delta) != delta) 1105 llvm_unreachable("Relocation R_PPC64_REL24 overflow"); 1106 // We preserve bits other than LI field, i.e. PO and AA/LK fields. 1107 uint32_t Inst = readBytesUnaligned(LocalAddress, 4); 1108 writeInt32BE(LocalAddress, (Inst & 0xFC000003) | (delta & 0x03FFFFFC)); 1109 } break; 1110 case ELF::R_PPC64_REL32: { 1111 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1112 int64_t delta = static_cast<int64_t>(Value - FinalAddress + Addend); 1113 if (SignExtend64<32>(delta) != delta) 1114 llvm_unreachable("Relocation R_PPC64_REL32 overflow"); 1115 writeInt32BE(LocalAddress, delta); 1116 } break; 1117 case ELF::R_PPC64_REL64: { 1118 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1119 uint64_t Delta = Value - FinalAddress + Addend; 1120 writeInt64BE(LocalAddress, Delta); 1121 } break; 1122 case ELF::R_PPC64_ADDR64: 1123 writeInt64BE(LocalAddress, Value + Addend); 1124 break; 1125 } 1126 } 1127 1128 void RuntimeDyldELF::resolveSystemZRelocation(const SectionEntry &Section, 1129 uint64_t Offset, uint64_t Value, 1130 uint32_t Type, int64_t Addend) { 1131 uint8_t *LocalAddress = Section.getAddressWithOffset(Offset); 1132 switch (Type) { 1133 default: 1134 report_fatal_error("Relocation type not implemented yet!"); 1135 break; 1136 case ELF::R_390_PC16DBL: 1137 case ELF::R_390_PLT16DBL: { 1138 int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(Offset); 1139 assert(int16_t(Delta / 2) * 2 == Delta && "R_390_PC16DBL overflow"); 1140 writeInt16BE(LocalAddress, Delta / 2); 1141 break; 1142 } 1143 case ELF::R_390_PC32DBL: 1144 case ELF::R_390_PLT32DBL: { 1145 int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(Offset); 1146 assert(int32_t(Delta / 2) * 2 == Delta && "R_390_PC32DBL overflow"); 1147 writeInt32BE(LocalAddress, Delta / 2); 1148 break; 1149 } 1150 case ELF::R_390_PC16: { 1151 int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(Offset); 1152 assert(int16_t(Delta) == Delta && "R_390_PC16 overflow"); 1153 writeInt16BE(LocalAddress, Delta); 1154 break; 1155 } 1156 case ELF::R_390_PC32: { 1157 int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(Offset); 1158 assert(int32_t(Delta) == Delta && "R_390_PC32 overflow"); 1159 writeInt32BE(LocalAddress, Delta); 1160 break; 1161 } 1162 case ELF::R_390_PC64: { 1163 int64_t Delta = (Value + Addend) - Section.getLoadAddressWithOffset(Offset); 1164 writeInt64BE(LocalAddress, Delta); 1165 break; 1166 } 1167 case ELF::R_390_8: 1168 *LocalAddress = (uint8_t)(Value + Addend); 1169 break; 1170 case ELF::R_390_16: 1171 writeInt16BE(LocalAddress, Value + Addend); 1172 break; 1173 case ELF::R_390_32: 1174 writeInt32BE(LocalAddress, Value + Addend); 1175 break; 1176 case ELF::R_390_64: 1177 writeInt64BE(LocalAddress, Value + Addend); 1178 break; 1179 } 1180 } 1181 1182 void RuntimeDyldELF::resolveBPFRelocation(const SectionEntry &Section, 1183 uint64_t Offset, uint64_t Value, 1184 uint32_t Type, int64_t Addend) { 1185 bool isBE = Arch == Triple::bpfeb; 1186 1187 switch (Type) { 1188 default: 1189 report_fatal_error("Relocation type not implemented yet!"); 1190 break; 1191 case ELF::R_BPF_NONE: 1192 case ELF::R_BPF_64_64: 1193 case ELF::R_BPF_64_32: 1194 case ELF::R_BPF_64_NODYLD32: 1195 break; 1196 case ELF::R_BPF_64_ABS64: { 1197 write(isBE, Section.getAddressWithOffset(Offset), Value + Addend); 1198 LLVM_DEBUG(dbgs() << "Writing " << format("%p", (Value + Addend)) << " at " 1199 << format("%p\n", Section.getAddressWithOffset(Offset))); 1200 break; 1201 } 1202 case ELF::R_BPF_64_ABS32: { 1203 Value += Addend; 1204 assert(Value <= UINT32_MAX); 1205 write(isBE, Section.getAddressWithOffset(Offset), static_cast<uint32_t>(Value)); 1206 LLVM_DEBUG(dbgs() << "Writing " << format("%p", Value) << " at " 1207 << format("%p\n", Section.getAddressWithOffset(Offset))); 1208 break; 1209 } 1210 } 1211 } 1212 1213 static void applyUTypeImmRISCV(uint8_t *InstrAddr, uint32_t Imm) { 1214 uint32_t UpperImm = (Imm + 0x800) & 0xfffff000; 1215 auto Instr = support::ulittle32_t::ref(InstrAddr); 1216 Instr = (Instr & 0xfff) | UpperImm; 1217 } 1218 1219 static void applyITypeImmRISCV(uint8_t *InstrAddr, uint32_t Imm) { 1220 uint32_t LowerImm = Imm & 0xfff; 1221 auto Instr = support::ulittle32_t::ref(InstrAddr); 1222 Instr = (Instr & 0xfffff) | (LowerImm << 20); 1223 } 1224 1225 void RuntimeDyldELF::resolveRISCVRelocation(const SectionEntry &Section, 1226 uint64_t Offset, uint64_t Value, 1227 uint32_t Type, int64_t Addend, 1228 SID SectionID) { 1229 switch (Type) { 1230 default: { 1231 std::string Err = "Unimplemented reloc type: " + std::to_string(Type); 1232 llvm::report_fatal_error(Err.c_str()); 1233 } 1234 // 32-bit PC-relative function call, macros call, tail (PIC) 1235 // Write first 20 bits of 32 bit value to the auipc instruction 1236 // Last 12 bits to the jalr instruction 1237 case ELF::R_RISCV_CALL: 1238 case ELF::R_RISCV_CALL_PLT: { 1239 uint64_t P = Section.getLoadAddressWithOffset(Offset); 1240 uint64_t PCOffset = Value + Addend - P; 1241 applyUTypeImmRISCV(Section.getAddressWithOffset(Offset), PCOffset); 1242 applyITypeImmRISCV(Section.getAddressWithOffset(Offset + 4), PCOffset); 1243 break; 1244 } 1245 // High 20 bits of 32-bit absolute address, %hi(symbol) 1246 case ELF::R_RISCV_HI20: { 1247 uint64_t PCOffset = Value + Addend; 1248 applyUTypeImmRISCV(Section.getAddressWithOffset(Offset), PCOffset); 1249 break; 1250 } 1251 // Low 12 bits of 32-bit absolute address, %lo(symbol) 1252 case ELF::R_RISCV_LO12_I: { 1253 uint64_t PCOffset = Value + Addend; 1254 applyITypeImmRISCV(Section.getAddressWithOffset(Offset), PCOffset); 1255 break; 1256 } 1257 // High 20 bits of 32-bit PC-relative reference, %pcrel_hi(symbol) 1258 case ELF::R_RISCV_GOT_HI20: 1259 case ELF::R_RISCV_PCREL_HI20: { 1260 uint64_t P = Section.getLoadAddressWithOffset(Offset); 1261 uint64_t PCOffset = Value + Addend - P; 1262 applyUTypeImmRISCV(Section.getAddressWithOffset(Offset), PCOffset); 1263 break; 1264 } 1265 1266 // label: 1267 // auipc a0, %pcrel_hi(symbol) // R_RISCV_PCREL_HI20 1268 // addi a0, a0, %pcrel_lo(label) // R_RISCV_PCREL_LO12_I 1269 // 1270 // The low 12 bits of relative address between pc and symbol. 1271 // The symbol is related to the high part instruction which is marked by 1272 // label. 1273 case ELF::R_RISCV_PCREL_LO12_I: { 1274 for (auto &&PendingReloc : PendingRelocs) { 1275 const RelocationValueRef &MatchingValue = PendingReloc.first; 1276 RelocationEntry &Reloc = PendingReloc.second; 1277 uint64_t HIRelocPC = 1278 getSectionLoadAddress(Reloc.SectionID) + Reloc.Offset; 1279 if (Value + Addend == HIRelocPC) { 1280 uint64_t Symbol = getSectionLoadAddress(MatchingValue.SectionID) + 1281 MatchingValue.Addend; 1282 auto PCOffset = Symbol - HIRelocPC; 1283 applyITypeImmRISCV(Section.getAddressWithOffset(Offset), PCOffset); 1284 return; 1285 } 1286 } 1287 1288 llvm::report_fatal_error( 1289 "R_RISCV_PCREL_LO12_I without matching R_RISCV_PCREL_HI20"); 1290 } 1291 case ELF::R_RISCV_32_PCREL: { 1292 uint64_t FinalAddress = Section.getLoadAddressWithOffset(Offset); 1293 int64_t RealOffset = Value + Addend - FinalAddress; 1294 int32_t TruncOffset = Lo_32(RealOffset); 1295 support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)) = 1296 TruncOffset; 1297 break; 1298 } 1299 case ELF::R_RISCV_32: { 1300 auto Ref = support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)); 1301 Ref = Value + Addend; 1302 break; 1303 } 1304 case ELF::R_RISCV_64: { 1305 auto Ref = support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)); 1306 Ref = Value + Addend; 1307 break; 1308 } 1309 case ELF::R_RISCV_ADD16: { 1310 auto Ref = support::ulittle16_t::ref(Section.getAddressWithOffset(Offset)); 1311 Ref = Ref + Value + Addend; 1312 break; 1313 } 1314 case ELF::R_RISCV_ADD32: { 1315 auto Ref = support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)); 1316 Ref = Ref + Value + Addend; 1317 break; 1318 } 1319 case ELF::R_RISCV_ADD64: { 1320 auto Ref = support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)); 1321 Ref = Ref + Value + Addend; 1322 break; 1323 } 1324 case ELF::R_RISCV_SUB16: { 1325 auto Ref = support::ulittle16_t::ref(Section.getAddressWithOffset(Offset)); 1326 Ref = Ref - Value - Addend; 1327 break; 1328 } 1329 case ELF::R_RISCV_SUB32: { 1330 auto Ref = support::ulittle32_t::ref(Section.getAddressWithOffset(Offset)); 1331 Ref = Ref - Value - Addend; 1332 break; 1333 } 1334 case ELF::R_RISCV_SUB64: { 1335 auto Ref = support::ulittle64_t::ref(Section.getAddressWithOffset(Offset)); 1336 Ref = Ref - Value - Addend; 1337 break; 1338 } 1339 } 1340 } 1341 1342 // The target location for the relocation is described by RE.SectionID and 1343 // RE.Offset. RE.SectionID can be used to find the SectionEntry. Each 1344 // SectionEntry has three members describing its location. 1345 // SectionEntry::Address is the address at which the section has been loaded 1346 // into memory in the current (host) process. SectionEntry::LoadAddress is the 1347 // address that the section will have in the target process. 1348 // SectionEntry::ObjAddress is the address of the bits for this section in the 1349 // original emitted object image (also in the current address space). 1350 // 1351 // Relocations will be applied as if the section were loaded at 1352 // SectionEntry::LoadAddress, but they will be applied at an address based 1353 // on SectionEntry::Address. SectionEntry::ObjAddress will be used to refer to 1354 // Target memory contents if they are required for value calculations. 1355 // 1356 // The Value parameter here is the load address of the symbol for the 1357 // relocation to be applied. For relocations which refer to symbols in the 1358 // current object Value will be the LoadAddress of the section in which 1359 // the symbol resides (RE.Addend provides additional information about the 1360 // symbol location). For external symbols, Value will be the address of the 1361 // symbol in the target address space. 1362 void RuntimeDyldELF::resolveRelocation(const RelocationEntry &RE, 1363 uint64_t Value) { 1364 const SectionEntry &Section = Sections[RE.SectionID]; 1365 return resolveRelocation(Section, RE.Offset, Value, RE.RelType, RE.Addend, 1366 RE.SymOffset, RE.SectionID); 1367 } 1368 1369 void RuntimeDyldELF::resolveRelocation(const SectionEntry &Section, 1370 uint64_t Offset, uint64_t Value, 1371 uint32_t Type, int64_t Addend, 1372 uint64_t SymOffset, SID SectionID) { 1373 switch (Arch) { 1374 case Triple::x86_64: 1375 resolveX86_64Relocation(Section, Offset, Value, Type, Addend, SymOffset); 1376 break; 1377 case Triple::x86: 1378 resolveX86Relocation(Section, Offset, (uint32_t)(Value & 0xffffffffL), Type, 1379 (uint32_t)(Addend & 0xffffffffL)); 1380 break; 1381 case Triple::aarch64: 1382 case Triple::aarch64_be: 1383 resolveAArch64Relocation(Section, Offset, Value, Type, Addend); 1384 break; 1385 case Triple::arm: // Fall through. 1386 case Triple::armeb: 1387 case Triple::thumb: 1388 case Triple::thumbeb: 1389 resolveARMRelocation(Section, Offset, (uint32_t)(Value & 0xffffffffL), Type, 1390 (uint32_t)(Addend & 0xffffffffL)); 1391 break; 1392 case Triple::loongarch64: 1393 resolveLoongArch64Relocation(Section, Offset, Value, Type, Addend); 1394 break; 1395 case Triple::ppc: // Fall through. 1396 case Triple::ppcle: 1397 resolvePPC32Relocation(Section, Offset, Value, Type, Addend); 1398 break; 1399 case Triple::ppc64: // Fall through. 1400 case Triple::ppc64le: 1401 resolvePPC64Relocation(Section, Offset, Value, Type, Addend); 1402 break; 1403 case Triple::systemz: 1404 resolveSystemZRelocation(Section, Offset, Value, Type, Addend); 1405 break; 1406 case Triple::bpfel: 1407 case Triple::bpfeb: 1408 resolveBPFRelocation(Section, Offset, Value, Type, Addend); 1409 break; 1410 case Triple::riscv32: // Fall through. 1411 case Triple::riscv64: 1412 resolveRISCVRelocation(Section, Offset, Value, Type, Addend, SectionID); 1413 break; 1414 default: 1415 llvm_unreachable("Unsupported CPU type!"); 1416 } 1417 } 1418 1419 void *RuntimeDyldELF::computePlaceholderAddress(unsigned SectionID, 1420 uint64_t Offset) const { 1421 return (void *)(Sections[SectionID].getObjAddress() + Offset); 1422 } 1423 1424 void RuntimeDyldELF::processSimpleRelocation(unsigned SectionID, uint64_t Offset, unsigned RelType, RelocationValueRef Value) { 1425 RelocationEntry RE(SectionID, Offset, RelType, Value.Addend, Value.Offset); 1426 if (Value.SymbolName) 1427 addRelocationForSymbol(RE, Value.SymbolName); 1428 else 1429 addRelocationForSection(RE, Value.SectionID); 1430 } 1431 1432 uint32_t RuntimeDyldELF::getMatchingLoRelocation(uint32_t RelType, 1433 bool IsLocal) const { 1434 switch (RelType) { 1435 case ELF::R_MICROMIPS_GOT16: 1436 if (IsLocal) 1437 return ELF::R_MICROMIPS_LO16; 1438 break; 1439 case ELF::R_MICROMIPS_HI16: 1440 return ELF::R_MICROMIPS_LO16; 1441 case ELF::R_MIPS_GOT16: 1442 if (IsLocal) 1443 return ELF::R_MIPS_LO16; 1444 break; 1445 case ELF::R_MIPS_HI16: 1446 return ELF::R_MIPS_LO16; 1447 case ELF::R_MIPS_PCHI16: 1448 return ELF::R_MIPS_PCLO16; 1449 default: 1450 break; 1451 } 1452 return ELF::R_MIPS_NONE; 1453 } 1454 1455 // Sometimes we don't need to create thunk for a branch. 1456 // This typically happens when branch target is located 1457 // in the same object file. In such case target is either 1458 // a weak symbol or symbol in a different executable section. 1459 // This function checks if branch target is located in the 1460 // same object file and if distance between source and target 1461 // fits R_AARCH64_CALL26 relocation. If both conditions are 1462 // met, it emits direct jump to the target and returns true. 1463 // Otherwise false is returned and thunk is created. 1464 bool RuntimeDyldELF::resolveAArch64ShortBranch( 1465 unsigned SectionID, relocation_iterator RelI, 1466 const RelocationValueRef &Value) { 1467 uint64_t TargetOffset; 1468 unsigned TargetSectionID; 1469 if (Value.SymbolName) { 1470 auto Loc = GlobalSymbolTable.find(Value.SymbolName); 1471 1472 // Don't create direct branch for external symbols. 1473 if (Loc == GlobalSymbolTable.end()) 1474 return false; 1475 1476 const auto &SymInfo = Loc->second; 1477 1478 TargetSectionID = SymInfo.getSectionID(); 1479 TargetOffset = SymInfo.getOffset(); 1480 } else { 1481 TargetSectionID = Value.SectionID; 1482 TargetOffset = 0; 1483 } 1484 1485 // We don't actually know the load addresses at this point, so if the 1486 // branch is cross-section, we don't know exactly how far away it is. 1487 if (TargetSectionID != SectionID) 1488 return false; 1489 1490 uint64_t SourceOffset = RelI->getOffset(); 1491 1492 // R_AARCH64_CALL26 requires immediate to be in range -2^27 <= imm < 2^27 1493 // If distance between source and target is out of range then we should 1494 // create thunk. 1495 if (!isInt<28>(TargetOffset + Value.Addend - SourceOffset)) 1496 return false; 1497 1498 RelocationEntry RE(SectionID, SourceOffset, RelI->getType(), Value.Addend); 1499 if (Value.SymbolName) 1500 addRelocationForSymbol(RE, Value.SymbolName); 1501 else 1502 addRelocationForSection(RE, Value.SectionID); 1503 1504 return true; 1505 } 1506 1507 void RuntimeDyldELF::resolveAArch64Branch(unsigned SectionID, 1508 const RelocationValueRef &Value, 1509 relocation_iterator RelI, 1510 StubMap &Stubs) { 1511 1512 LLVM_DEBUG(dbgs() << "\t\tThis is an AArch64 branch relocation."); 1513 SectionEntry &Section = Sections[SectionID]; 1514 1515 uint64_t Offset = RelI->getOffset(); 1516 unsigned RelType = RelI->getType(); 1517 // Look for an existing stub. 1518 StubMap::const_iterator i = Stubs.find(Value); 1519 if (i != Stubs.end()) { 1520 resolveRelocation(Section, Offset, 1521 Section.getLoadAddressWithOffset(i->second), RelType, 0); 1522 LLVM_DEBUG(dbgs() << " Stub function found\n"); 1523 } else if (!resolveAArch64ShortBranch(SectionID, RelI, Value)) { 1524 // Create a new stub function. 1525 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 1526 Stubs[Value] = Section.getStubOffset(); 1527 uint8_t *StubTargetAddr = createStubFunction( 1528 Section.getAddressWithOffset(Section.getStubOffset())); 1529 1530 RelocationEntry REmovz_g3(SectionID, StubTargetAddr - Section.getAddress(), 1531 ELF::R_AARCH64_MOVW_UABS_G3, Value.Addend); 1532 RelocationEntry REmovk_g2(SectionID, 1533 StubTargetAddr - Section.getAddress() + 4, 1534 ELF::R_AARCH64_MOVW_UABS_G2_NC, Value.Addend); 1535 RelocationEntry REmovk_g1(SectionID, 1536 StubTargetAddr - Section.getAddress() + 8, 1537 ELF::R_AARCH64_MOVW_UABS_G1_NC, Value.Addend); 1538 RelocationEntry REmovk_g0(SectionID, 1539 StubTargetAddr - Section.getAddress() + 12, 1540 ELF::R_AARCH64_MOVW_UABS_G0_NC, Value.Addend); 1541 1542 if (Value.SymbolName) { 1543 addRelocationForSymbol(REmovz_g3, Value.SymbolName); 1544 addRelocationForSymbol(REmovk_g2, Value.SymbolName); 1545 addRelocationForSymbol(REmovk_g1, Value.SymbolName); 1546 addRelocationForSymbol(REmovk_g0, Value.SymbolName); 1547 } else { 1548 addRelocationForSection(REmovz_g3, Value.SectionID); 1549 addRelocationForSection(REmovk_g2, Value.SectionID); 1550 addRelocationForSection(REmovk_g1, Value.SectionID); 1551 addRelocationForSection(REmovk_g0, Value.SectionID); 1552 } 1553 resolveRelocation(Section, Offset, 1554 Section.getLoadAddressWithOffset(Section.getStubOffset()), 1555 RelType, 0); 1556 Section.advanceStubOffset(getMaxStubSize()); 1557 } 1558 } 1559 1560 Expected<relocation_iterator> 1561 RuntimeDyldELF::processRelocationRef( 1562 unsigned SectionID, relocation_iterator RelI, const ObjectFile &O, 1563 ObjSectionToIDMap &ObjSectionToID, StubMap &Stubs) { 1564 const auto &Obj = cast<ELFObjectFileBase>(O); 1565 uint64_t RelType = RelI->getType(); 1566 int64_t Addend = 0; 1567 if (Expected<int64_t> AddendOrErr = ELFRelocationRef(*RelI).getAddend()) 1568 Addend = *AddendOrErr; 1569 else 1570 consumeError(AddendOrErr.takeError()); 1571 elf_symbol_iterator Symbol = RelI->getSymbol(); 1572 1573 // Obtain the symbol name which is referenced in the relocation 1574 StringRef TargetName; 1575 if (Symbol != Obj.symbol_end()) { 1576 if (auto TargetNameOrErr = Symbol->getName()) 1577 TargetName = *TargetNameOrErr; 1578 else 1579 return TargetNameOrErr.takeError(); 1580 } 1581 LLVM_DEBUG(dbgs() << "\t\tRelType: " << RelType << " Addend: " << Addend 1582 << " TargetName: " << TargetName << "\n"); 1583 RelocationValueRef Value; 1584 // First search for the symbol in the local symbol table 1585 SymbolRef::Type SymType = SymbolRef::ST_Unknown; 1586 1587 // Search for the symbol in the global symbol table 1588 RTDyldSymbolTable::const_iterator gsi = GlobalSymbolTable.end(); 1589 if (Symbol != Obj.symbol_end()) { 1590 gsi = GlobalSymbolTable.find(TargetName.data()); 1591 Expected<SymbolRef::Type> SymTypeOrErr = Symbol->getType(); 1592 if (!SymTypeOrErr) { 1593 std::string Buf; 1594 raw_string_ostream OS(Buf); 1595 logAllUnhandledErrors(SymTypeOrErr.takeError(), OS); 1596 report_fatal_error(Twine(Buf)); 1597 } 1598 SymType = *SymTypeOrErr; 1599 } 1600 if (gsi != GlobalSymbolTable.end()) { 1601 const auto &SymInfo = gsi->second; 1602 Value.SectionID = SymInfo.getSectionID(); 1603 Value.Offset = SymInfo.getOffset(); 1604 Value.Addend = SymInfo.getOffset() + Addend; 1605 } else { 1606 switch (SymType) { 1607 case SymbolRef::ST_Debug: { 1608 // TODO: Now ELF SymbolRef::ST_Debug = STT_SECTION, it's not obviously 1609 // and can be changed by another developers. Maybe best way is add 1610 // a new symbol type ST_Section to SymbolRef and use it. 1611 auto SectionOrErr = Symbol->getSection(); 1612 if (!SectionOrErr) { 1613 std::string Buf; 1614 raw_string_ostream OS(Buf); 1615 logAllUnhandledErrors(SectionOrErr.takeError(), OS); 1616 report_fatal_error(Twine(Buf)); 1617 } 1618 section_iterator si = *SectionOrErr; 1619 if (si == Obj.section_end()) 1620 llvm_unreachable("Symbol section not found, bad object file format!"); 1621 LLVM_DEBUG(dbgs() << "\t\tThis is section symbol\n"); 1622 bool isCode = si->isText(); 1623 if (auto SectionIDOrErr = findOrEmitSection(Obj, (*si), isCode, 1624 ObjSectionToID)) 1625 Value.SectionID = *SectionIDOrErr; 1626 else 1627 return SectionIDOrErr.takeError(); 1628 Value.Addend = Addend; 1629 break; 1630 } 1631 case SymbolRef::ST_Data: 1632 case SymbolRef::ST_Function: 1633 case SymbolRef::ST_Other: 1634 case SymbolRef::ST_Unknown: { 1635 Value.SymbolName = TargetName.data(); 1636 Value.Addend = Addend; 1637 1638 // Absolute relocations will have a zero symbol ID (STN_UNDEF), which 1639 // will manifest here as a NULL symbol name. 1640 // We can set this as a valid (but empty) symbol name, and rely 1641 // on addRelocationForSymbol to handle this. 1642 if (!Value.SymbolName) 1643 Value.SymbolName = ""; 1644 break; 1645 } 1646 default: 1647 llvm_unreachable("Unresolved symbol type!"); 1648 break; 1649 } 1650 } 1651 1652 uint64_t Offset = RelI->getOffset(); 1653 1654 LLVM_DEBUG(dbgs() << "\t\tSectionID: " << SectionID << " Offset: " << Offset 1655 << "\n"); 1656 if ((Arch == Triple::aarch64 || Arch == Triple::aarch64_be)) { 1657 if ((RelType == ELF::R_AARCH64_CALL26 || 1658 RelType == ELF::R_AARCH64_JUMP26) && 1659 MemMgr.allowStubAllocation()) { 1660 resolveAArch64Branch(SectionID, Value, RelI, Stubs); 1661 } else if (RelType == ELF::R_AARCH64_ADR_GOT_PAGE) { 1662 // Create new GOT entry or find existing one. If GOT entry is 1663 // to be created, then we also emit ABS64 relocation for it. 1664 uint64_t GOTOffset = findOrAllocGOTEntry(Value, ELF::R_AARCH64_ABS64); 1665 resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend, 1666 ELF::R_AARCH64_ADR_PREL_PG_HI21); 1667 1668 } else if (RelType == ELF::R_AARCH64_LD64_GOT_LO12_NC) { 1669 uint64_t GOTOffset = findOrAllocGOTEntry(Value, ELF::R_AARCH64_ABS64); 1670 resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend, 1671 ELF::R_AARCH64_LDST64_ABS_LO12_NC); 1672 } else { 1673 processSimpleRelocation(SectionID, Offset, RelType, Value); 1674 } 1675 } else if (Arch == Triple::arm) { 1676 if (RelType == ELF::R_ARM_PC24 || RelType == ELF::R_ARM_CALL || 1677 RelType == ELF::R_ARM_JUMP24) { 1678 // This is an ARM branch relocation, need to use a stub function. 1679 LLVM_DEBUG(dbgs() << "\t\tThis is an ARM branch relocation.\n"); 1680 SectionEntry &Section = Sections[SectionID]; 1681 1682 // Look for an existing stub. 1683 StubMap::const_iterator i = Stubs.find(Value); 1684 if (i != Stubs.end()) { 1685 resolveRelocation(Section, Offset, 1686 Section.getLoadAddressWithOffset(i->second), RelType, 1687 0); 1688 LLVM_DEBUG(dbgs() << " Stub function found\n"); 1689 } else { 1690 // Create a new stub function. 1691 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 1692 Stubs[Value] = Section.getStubOffset(); 1693 uint8_t *StubTargetAddr = createStubFunction( 1694 Section.getAddressWithOffset(Section.getStubOffset())); 1695 RelocationEntry RE(SectionID, StubTargetAddr - Section.getAddress(), 1696 ELF::R_ARM_ABS32, Value.Addend); 1697 if (Value.SymbolName) 1698 addRelocationForSymbol(RE, Value.SymbolName); 1699 else 1700 addRelocationForSection(RE, Value.SectionID); 1701 1702 resolveRelocation( 1703 Section, Offset, 1704 Section.getLoadAddressWithOffset(Section.getStubOffset()), RelType, 1705 0); 1706 Section.advanceStubOffset(getMaxStubSize()); 1707 } 1708 } else { 1709 uint32_t *Placeholder = 1710 reinterpret_cast<uint32_t*>(computePlaceholderAddress(SectionID, Offset)); 1711 if (RelType == ELF::R_ARM_PREL31 || RelType == ELF::R_ARM_TARGET1 || 1712 RelType == ELF::R_ARM_ABS32) { 1713 Value.Addend += *Placeholder; 1714 } else if (RelType == ELF::R_ARM_MOVW_ABS_NC || RelType == ELF::R_ARM_MOVT_ABS) { 1715 // See ELF for ARM documentation 1716 Value.Addend += (int16_t)((*Placeholder & 0xFFF) | (((*Placeholder >> 16) & 0xF) << 12)); 1717 } 1718 processSimpleRelocation(SectionID, Offset, RelType, Value); 1719 } 1720 } else if (Arch == Triple::loongarch64) { 1721 if (RelType == ELF::R_LARCH_B26 && MemMgr.allowStubAllocation()) { 1722 resolveLoongArch64Branch(SectionID, Value, RelI, Stubs); 1723 } else if (RelType == ELF::R_LARCH_GOT_PC_HI20 || 1724 RelType == ELF::R_LARCH_GOT_PC_LO12) { 1725 uint64_t GOTOffset = findOrAllocGOTEntry(Value, ELF::R_LARCH_64); 1726 resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend, 1727 RelType); 1728 } else { 1729 processSimpleRelocation(SectionID, Offset, RelType, Value); 1730 } 1731 } else if (IsMipsO32ABI) { 1732 uint8_t *Placeholder = reinterpret_cast<uint8_t *>( 1733 computePlaceholderAddress(SectionID, Offset)); 1734 uint32_t Opcode = readBytesUnaligned(Placeholder, 4); 1735 if (RelType == ELF::R_MIPS_26) { 1736 // This is an Mips branch relocation, need to use a stub function. 1737 LLVM_DEBUG(dbgs() << "\t\tThis is a Mips branch relocation."); 1738 SectionEntry &Section = Sections[SectionID]; 1739 1740 // Extract the addend from the instruction. 1741 // We shift up by two since the Value will be down shifted again 1742 // when applying the relocation. 1743 uint32_t Addend = (Opcode & 0x03ffffff) << 2; 1744 1745 Value.Addend += Addend; 1746 1747 // Look up for existing stub. 1748 StubMap::const_iterator i = Stubs.find(Value); 1749 if (i != Stubs.end()) { 1750 RelocationEntry RE(SectionID, Offset, RelType, i->second); 1751 addRelocationForSection(RE, SectionID); 1752 LLVM_DEBUG(dbgs() << " Stub function found\n"); 1753 } else { 1754 // Create a new stub function. 1755 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 1756 Stubs[Value] = Section.getStubOffset(); 1757 1758 unsigned AbiVariant = Obj.getPlatformFlags(); 1759 1760 uint8_t *StubTargetAddr = createStubFunction( 1761 Section.getAddressWithOffset(Section.getStubOffset()), AbiVariant); 1762 1763 // Creating Hi and Lo relocations for the filled stub instructions. 1764 RelocationEntry REHi(SectionID, StubTargetAddr - Section.getAddress(), 1765 ELF::R_MIPS_HI16, Value.Addend); 1766 RelocationEntry RELo(SectionID, 1767 StubTargetAddr - Section.getAddress() + 4, 1768 ELF::R_MIPS_LO16, Value.Addend); 1769 1770 if (Value.SymbolName) { 1771 addRelocationForSymbol(REHi, Value.SymbolName); 1772 addRelocationForSymbol(RELo, Value.SymbolName); 1773 } else { 1774 addRelocationForSection(REHi, Value.SectionID); 1775 addRelocationForSection(RELo, Value.SectionID); 1776 } 1777 1778 RelocationEntry RE(SectionID, Offset, RelType, Section.getStubOffset()); 1779 addRelocationForSection(RE, SectionID); 1780 Section.advanceStubOffset(getMaxStubSize()); 1781 } 1782 } else if (RelType == ELF::R_MIPS_HI16 || RelType == ELF::R_MIPS_PCHI16) { 1783 int64_t Addend = (Opcode & 0x0000ffff) << 16; 1784 RelocationEntry RE(SectionID, Offset, RelType, Addend); 1785 PendingRelocs.push_back(std::make_pair(Value, RE)); 1786 } else if (RelType == ELF::R_MIPS_LO16 || RelType == ELF::R_MIPS_PCLO16) { 1787 int64_t Addend = Value.Addend + SignExtend32<16>(Opcode & 0x0000ffff); 1788 for (auto I = PendingRelocs.begin(); I != PendingRelocs.end();) { 1789 const RelocationValueRef &MatchingValue = I->first; 1790 RelocationEntry &Reloc = I->second; 1791 if (MatchingValue == Value && 1792 RelType == getMatchingLoRelocation(Reloc.RelType) && 1793 SectionID == Reloc.SectionID) { 1794 Reloc.Addend += Addend; 1795 if (Value.SymbolName) 1796 addRelocationForSymbol(Reloc, Value.SymbolName); 1797 else 1798 addRelocationForSection(Reloc, Value.SectionID); 1799 I = PendingRelocs.erase(I); 1800 } else 1801 ++I; 1802 } 1803 RelocationEntry RE(SectionID, Offset, RelType, Addend); 1804 if (Value.SymbolName) 1805 addRelocationForSymbol(RE, Value.SymbolName); 1806 else 1807 addRelocationForSection(RE, Value.SectionID); 1808 } else { 1809 if (RelType == ELF::R_MIPS_32) 1810 Value.Addend += Opcode; 1811 else if (RelType == ELF::R_MIPS_PC16) 1812 Value.Addend += SignExtend32<18>((Opcode & 0x0000ffff) << 2); 1813 else if (RelType == ELF::R_MIPS_PC19_S2) 1814 Value.Addend += SignExtend32<21>((Opcode & 0x0007ffff) << 2); 1815 else if (RelType == ELF::R_MIPS_PC21_S2) 1816 Value.Addend += SignExtend32<23>((Opcode & 0x001fffff) << 2); 1817 else if (RelType == ELF::R_MIPS_PC26_S2) 1818 Value.Addend += SignExtend32<28>((Opcode & 0x03ffffff) << 2); 1819 processSimpleRelocation(SectionID, Offset, RelType, Value); 1820 } 1821 } else if (IsMipsN32ABI || IsMipsN64ABI) { 1822 uint32_t r_type = RelType & 0xff; 1823 RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); 1824 if (r_type == ELF::R_MIPS_CALL16 || r_type == ELF::R_MIPS_GOT_PAGE 1825 || r_type == ELF::R_MIPS_GOT_DISP) { 1826 auto [I, Inserted] = GOTSymbolOffsets.try_emplace(TargetName); 1827 if (Inserted) 1828 I->second = allocateGOTEntries(1); 1829 RE.SymOffset = I->second; 1830 if (Value.SymbolName) 1831 addRelocationForSymbol(RE, Value.SymbolName); 1832 else 1833 addRelocationForSection(RE, Value.SectionID); 1834 } else if (RelType == ELF::R_MIPS_26) { 1835 // This is an Mips branch relocation, need to use a stub function. 1836 LLVM_DEBUG(dbgs() << "\t\tThis is a Mips branch relocation."); 1837 SectionEntry &Section = Sections[SectionID]; 1838 1839 // Look up for existing stub. 1840 StubMap::const_iterator i = Stubs.find(Value); 1841 if (i != Stubs.end()) { 1842 RelocationEntry RE(SectionID, Offset, RelType, i->second); 1843 addRelocationForSection(RE, SectionID); 1844 LLVM_DEBUG(dbgs() << " Stub function found\n"); 1845 } else { 1846 // Create a new stub function. 1847 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 1848 Stubs[Value] = Section.getStubOffset(); 1849 1850 unsigned AbiVariant = Obj.getPlatformFlags(); 1851 1852 uint8_t *StubTargetAddr = createStubFunction( 1853 Section.getAddressWithOffset(Section.getStubOffset()), AbiVariant); 1854 1855 if (IsMipsN32ABI) { 1856 // Creating Hi and Lo relocations for the filled stub instructions. 1857 RelocationEntry REHi(SectionID, StubTargetAddr - Section.getAddress(), 1858 ELF::R_MIPS_HI16, Value.Addend); 1859 RelocationEntry RELo(SectionID, 1860 StubTargetAddr - Section.getAddress() + 4, 1861 ELF::R_MIPS_LO16, Value.Addend); 1862 if (Value.SymbolName) { 1863 addRelocationForSymbol(REHi, Value.SymbolName); 1864 addRelocationForSymbol(RELo, Value.SymbolName); 1865 } else { 1866 addRelocationForSection(REHi, Value.SectionID); 1867 addRelocationForSection(RELo, Value.SectionID); 1868 } 1869 } else { 1870 // Creating Highest, Higher, Hi and Lo relocations for the filled stub 1871 // instructions. 1872 RelocationEntry REHighest(SectionID, 1873 StubTargetAddr - Section.getAddress(), 1874 ELF::R_MIPS_HIGHEST, Value.Addend); 1875 RelocationEntry REHigher(SectionID, 1876 StubTargetAddr - Section.getAddress() + 4, 1877 ELF::R_MIPS_HIGHER, Value.Addend); 1878 RelocationEntry REHi(SectionID, 1879 StubTargetAddr - Section.getAddress() + 12, 1880 ELF::R_MIPS_HI16, Value.Addend); 1881 RelocationEntry RELo(SectionID, 1882 StubTargetAddr - Section.getAddress() + 20, 1883 ELF::R_MIPS_LO16, Value.Addend); 1884 if (Value.SymbolName) { 1885 addRelocationForSymbol(REHighest, Value.SymbolName); 1886 addRelocationForSymbol(REHigher, Value.SymbolName); 1887 addRelocationForSymbol(REHi, Value.SymbolName); 1888 addRelocationForSymbol(RELo, Value.SymbolName); 1889 } else { 1890 addRelocationForSection(REHighest, Value.SectionID); 1891 addRelocationForSection(REHigher, Value.SectionID); 1892 addRelocationForSection(REHi, Value.SectionID); 1893 addRelocationForSection(RELo, Value.SectionID); 1894 } 1895 } 1896 RelocationEntry RE(SectionID, Offset, RelType, Section.getStubOffset()); 1897 addRelocationForSection(RE, SectionID); 1898 Section.advanceStubOffset(getMaxStubSize()); 1899 } 1900 } else { 1901 processSimpleRelocation(SectionID, Offset, RelType, Value); 1902 } 1903 1904 } else if (Arch == Triple::ppc64 || Arch == Triple::ppc64le) { 1905 if (RelType == ELF::R_PPC64_REL24) { 1906 // Determine ABI variant in use for this object. 1907 unsigned AbiVariant = Obj.getPlatformFlags(); 1908 AbiVariant &= ELF::EF_PPC64_ABI; 1909 // A PPC branch relocation will need a stub function if the target is 1910 // an external symbol (either Value.SymbolName is set, or SymType is 1911 // Symbol::ST_Unknown) or if the target address is not within the 1912 // signed 24-bits branch address. 1913 SectionEntry &Section = Sections[SectionID]; 1914 uint8_t *Target = Section.getAddressWithOffset(Offset); 1915 bool RangeOverflow = false; 1916 bool IsExtern = Value.SymbolName || SymType == SymbolRef::ST_Unknown; 1917 if (!IsExtern) { 1918 if (AbiVariant != 2) { 1919 // In the ELFv1 ABI, a function call may point to the .opd entry, 1920 // so the final symbol value is calculated based on the relocation 1921 // values in the .opd section. 1922 if (auto Err = findOPDEntrySection(Obj, ObjSectionToID, Value)) 1923 return std::move(Err); 1924 } else { 1925 // In the ELFv2 ABI, a function symbol may provide a local entry 1926 // point, which must be used for direct calls. 1927 if (Value.SectionID == SectionID){ 1928 uint8_t SymOther = Symbol->getOther(); 1929 Value.Addend += ELF::decodePPC64LocalEntryOffset(SymOther); 1930 } 1931 } 1932 uint8_t *RelocTarget = 1933 Sections[Value.SectionID].getAddressWithOffset(Value.Addend); 1934 int64_t delta = static_cast<int64_t>(Target - RelocTarget); 1935 // If it is within 26-bits branch range, just set the branch target 1936 if (SignExtend64<26>(delta) != delta) { 1937 RangeOverflow = true; 1938 } else if ((AbiVariant != 2) || 1939 (AbiVariant == 2 && Value.SectionID == SectionID)) { 1940 RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); 1941 addRelocationForSection(RE, Value.SectionID); 1942 } 1943 } 1944 if (IsExtern || (AbiVariant == 2 && Value.SectionID != SectionID) || 1945 RangeOverflow) { 1946 // It is an external symbol (either Value.SymbolName is set, or 1947 // SymType is SymbolRef::ST_Unknown) or out of range. 1948 StubMap::const_iterator i = Stubs.find(Value); 1949 if (i != Stubs.end()) { 1950 // Symbol function stub already created, just relocate to it 1951 resolveRelocation(Section, Offset, 1952 Section.getLoadAddressWithOffset(i->second), 1953 RelType, 0); 1954 LLVM_DEBUG(dbgs() << " Stub function found\n"); 1955 } else { 1956 // Create a new stub function. 1957 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 1958 Stubs[Value] = Section.getStubOffset(); 1959 uint8_t *StubTargetAddr = createStubFunction( 1960 Section.getAddressWithOffset(Section.getStubOffset()), 1961 AbiVariant); 1962 RelocationEntry RE(SectionID, StubTargetAddr - Section.getAddress(), 1963 ELF::R_PPC64_ADDR64, Value.Addend); 1964 1965 // Generates the 64-bits address loads as exemplified in section 1966 // 4.5.1 in PPC64 ELF ABI. Note that the relocations need to 1967 // apply to the low part of the instructions, so we have to update 1968 // the offset according to the target endianness. 1969 uint64_t StubRelocOffset = StubTargetAddr - Section.getAddress(); 1970 if (!IsTargetLittleEndian) 1971 StubRelocOffset += 2; 1972 1973 RelocationEntry REhst(SectionID, StubRelocOffset + 0, 1974 ELF::R_PPC64_ADDR16_HIGHEST, Value.Addend); 1975 RelocationEntry REhr(SectionID, StubRelocOffset + 4, 1976 ELF::R_PPC64_ADDR16_HIGHER, Value.Addend); 1977 RelocationEntry REh(SectionID, StubRelocOffset + 12, 1978 ELF::R_PPC64_ADDR16_HI, Value.Addend); 1979 RelocationEntry REl(SectionID, StubRelocOffset + 16, 1980 ELF::R_PPC64_ADDR16_LO, Value.Addend); 1981 1982 if (Value.SymbolName) { 1983 addRelocationForSymbol(REhst, Value.SymbolName); 1984 addRelocationForSymbol(REhr, Value.SymbolName); 1985 addRelocationForSymbol(REh, Value.SymbolName); 1986 addRelocationForSymbol(REl, Value.SymbolName); 1987 } else { 1988 addRelocationForSection(REhst, Value.SectionID); 1989 addRelocationForSection(REhr, Value.SectionID); 1990 addRelocationForSection(REh, Value.SectionID); 1991 addRelocationForSection(REl, Value.SectionID); 1992 } 1993 1994 resolveRelocation( 1995 Section, Offset, 1996 Section.getLoadAddressWithOffset(Section.getStubOffset()), 1997 RelType, 0); 1998 Section.advanceStubOffset(getMaxStubSize()); 1999 } 2000 if (IsExtern || (AbiVariant == 2 && Value.SectionID != SectionID)) { 2001 // Restore the TOC for external calls 2002 if (AbiVariant == 2) 2003 writeInt32BE(Target + 4, 0xE8410018); // ld r2,24(r1) 2004 else 2005 writeInt32BE(Target + 4, 0xE8410028); // ld r2,40(r1) 2006 } 2007 } 2008 } else if (RelType == ELF::R_PPC64_TOC16 || 2009 RelType == ELF::R_PPC64_TOC16_DS || 2010 RelType == ELF::R_PPC64_TOC16_LO || 2011 RelType == ELF::R_PPC64_TOC16_LO_DS || 2012 RelType == ELF::R_PPC64_TOC16_HI || 2013 RelType == ELF::R_PPC64_TOC16_HA) { 2014 // These relocations are supposed to subtract the TOC address from 2015 // the final value. This does not fit cleanly into the RuntimeDyld 2016 // scheme, since there may be *two* sections involved in determining 2017 // the relocation value (the section of the symbol referred to by the 2018 // relocation, and the TOC section associated with the current module). 2019 // 2020 // Fortunately, these relocations are currently only ever generated 2021 // referring to symbols that themselves reside in the TOC, which means 2022 // that the two sections are actually the same. Thus they cancel out 2023 // and we can immediately resolve the relocation right now. 2024 switch (RelType) { 2025 case ELF::R_PPC64_TOC16: RelType = ELF::R_PPC64_ADDR16; break; 2026 case ELF::R_PPC64_TOC16_DS: RelType = ELF::R_PPC64_ADDR16_DS; break; 2027 case ELF::R_PPC64_TOC16_LO: RelType = ELF::R_PPC64_ADDR16_LO; break; 2028 case ELF::R_PPC64_TOC16_LO_DS: RelType = ELF::R_PPC64_ADDR16_LO_DS; break; 2029 case ELF::R_PPC64_TOC16_HI: RelType = ELF::R_PPC64_ADDR16_HI; break; 2030 case ELF::R_PPC64_TOC16_HA: RelType = ELF::R_PPC64_ADDR16_HA; break; 2031 default: llvm_unreachable("Wrong relocation type."); 2032 } 2033 2034 RelocationValueRef TOCValue; 2035 if (auto Err = findPPC64TOCSection(Obj, ObjSectionToID, TOCValue)) 2036 return std::move(Err); 2037 if (Value.SymbolName || Value.SectionID != TOCValue.SectionID) 2038 llvm_unreachable("Unsupported TOC relocation."); 2039 Value.Addend -= TOCValue.Addend; 2040 resolveRelocation(Sections[SectionID], Offset, Value.Addend, RelType, 0); 2041 } else { 2042 // There are two ways to refer to the TOC address directly: either 2043 // via a ELF::R_PPC64_TOC relocation (where both symbol and addend are 2044 // ignored), or via any relocation that refers to the magic ".TOC." 2045 // symbols (in which case the addend is respected). 2046 if (RelType == ELF::R_PPC64_TOC) { 2047 RelType = ELF::R_PPC64_ADDR64; 2048 if (auto Err = findPPC64TOCSection(Obj, ObjSectionToID, Value)) 2049 return std::move(Err); 2050 } else if (TargetName == ".TOC.") { 2051 if (auto Err = findPPC64TOCSection(Obj, ObjSectionToID, Value)) 2052 return std::move(Err); 2053 Value.Addend += Addend; 2054 } 2055 2056 RelocationEntry RE(SectionID, Offset, RelType, Value.Addend); 2057 2058 if (Value.SymbolName) 2059 addRelocationForSymbol(RE, Value.SymbolName); 2060 else 2061 addRelocationForSection(RE, Value.SectionID); 2062 } 2063 } else if (Arch == Triple::systemz && 2064 (RelType == ELF::R_390_PLT32DBL || RelType == ELF::R_390_GOTENT)) { 2065 // Create function stubs for both PLT and GOT references, regardless of 2066 // whether the GOT reference is to data or code. The stub contains the 2067 // full address of the symbol, as needed by GOT references, and the 2068 // executable part only adds an overhead of 8 bytes. 2069 // 2070 // We could try to conserve space by allocating the code and data 2071 // parts of the stub separately. However, as things stand, we allocate 2072 // a stub for every relocation, so using a GOT in JIT code should be 2073 // no less space efficient than using an explicit constant pool. 2074 LLVM_DEBUG(dbgs() << "\t\tThis is a SystemZ indirect relocation."); 2075 SectionEntry &Section = Sections[SectionID]; 2076 2077 // Look for an existing stub. 2078 StubMap::const_iterator i = Stubs.find(Value); 2079 uintptr_t StubAddress; 2080 if (i != Stubs.end()) { 2081 StubAddress = uintptr_t(Section.getAddressWithOffset(i->second)); 2082 LLVM_DEBUG(dbgs() << " Stub function found\n"); 2083 } else { 2084 // Create a new stub function. 2085 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 2086 2087 uintptr_t BaseAddress = uintptr_t(Section.getAddress()); 2088 StubAddress = 2089 alignTo(BaseAddress + Section.getStubOffset(), getStubAlignment()); 2090 unsigned StubOffset = StubAddress - BaseAddress; 2091 2092 Stubs[Value] = StubOffset; 2093 createStubFunction((uint8_t *)StubAddress); 2094 RelocationEntry RE(SectionID, StubOffset + 8, ELF::R_390_64, 2095 Value.Offset); 2096 if (Value.SymbolName) 2097 addRelocationForSymbol(RE, Value.SymbolName); 2098 else 2099 addRelocationForSection(RE, Value.SectionID); 2100 Section.advanceStubOffset(getMaxStubSize()); 2101 } 2102 2103 if (RelType == ELF::R_390_GOTENT) 2104 resolveRelocation(Section, Offset, StubAddress + 8, ELF::R_390_PC32DBL, 2105 Addend); 2106 else 2107 resolveRelocation(Section, Offset, StubAddress, RelType, Addend); 2108 } else if (Arch == Triple::x86_64) { 2109 if (RelType == ELF::R_X86_64_PLT32) { 2110 // The way the PLT relocations normally work is that the linker allocates 2111 // the 2112 // PLT and this relocation makes a PC-relative call into the PLT. The PLT 2113 // entry will then jump to an address provided by the GOT. On first call, 2114 // the 2115 // GOT address will point back into PLT code that resolves the symbol. After 2116 // the first call, the GOT entry points to the actual function. 2117 // 2118 // For local functions we're ignoring all of that here and just replacing 2119 // the PLT32 relocation type with PC32, which will translate the relocation 2120 // into a PC-relative call directly to the function. For external symbols we 2121 // can't be sure the function will be within 2^32 bytes of the call site, so 2122 // we need to create a stub, which calls into the GOT. This case is 2123 // equivalent to the usual PLT implementation except that we use the stub 2124 // mechanism in RuntimeDyld (which puts stubs at the end of the section) 2125 // rather than allocating a PLT section. 2126 if (Value.SymbolName && MemMgr.allowStubAllocation()) { 2127 // This is a call to an external function. 2128 // Look for an existing stub. 2129 SectionEntry *Section = &Sections[SectionID]; 2130 StubMap::const_iterator i = Stubs.find(Value); 2131 uintptr_t StubAddress; 2132 if (i != Stubs.end()) { 2133 StubAddress = uintptr_t(Section->getAddress()) + i->second; 2134 LLVM_DEBUG(dbgs() << " Stub function found\n"); 2135 } else { 2136 // Create a new stub function (equivalent to a PLT entry). 2137 LLVM_DEBUG(dbgs() << " Create a new stub function\n"); 2138 2139 uintptr_t BaseAddress = uintptr_t(Section->getAddress()); 2140 StubAddress = alignTo(BaseAddress + Section->getStubOffset(), 2141 getStubAlignment()); 2142 unsigned StubOffset = StubAddress - BaseAddress; 2143 Stubs[Value] = StubOffset; 2144 createStubFunction((uint8_t *)StubAddress); 2145 2146 // Bump our stub offset counter 2147 Section->advanceStubOffset(getMaxStubSize()); 2148 2149 // Allocate a GOT Entry 2150 uint64_t GOTOffset = allocateGOTEntries(1); 2151 // This potentially creates a new Section which potentially 2152 // invalidates the Section pointer, so reload it. 2153 Section = &Sections[SectionID]; 2154 2155 // The load of the GOT address has an addend of -4 2156 resolveGOTOffsetRelocation(SectionID, StubOffset + 2, GOTOffset - 4, 2157 ELF::R_X86_64_PC32); 2158 2159 // Fill in the value of the symbol we're targeting into the GOT 2160 addRelocationForSymbol( 2161 computeGOTOffsetRE(GOTOffset, 0, ELF::R_X86_64_64), 2162 Value.SymbolName); 2163 } 2164 2165 // Make the target call a call into the stub table. 2166 resolveRelocation(*Section, Offset, StubAddress, ELF::R_X86_64_PC32, 2167 Addend); 2168 } else { 2169 Value.Addend += support::ulittle32_t::ref( 2170 computePlaceholderAddress(SectionID, Offset)); 2171 processSimpleRelocation(SectionID, Offset, ELF::R_X86_64_PC32, Value); 2172 } 2173 } else if (RelType == ELF::R_X86_64_GOTPCREL || 2174 RelType == ELF::R_X86_64_GOTPCRELX || 2175 RelType == ELF::R_X86_64_REX_GOTPCRELX) { 2176 uint64_t GOTOffset = allocateGOTEntries(1); 2177 resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend, 2178 ELF::R_X86_64_PC32); 2179 2180 // Fill in the value of the symbol we're targeting into the GOT 2181 RelocationEntry RE = 2182 computeGOTOffsetRE(GOTOffset, Value.Offset, ELF::R_X86_64_64); 2183 if (Value.SymbolName) 2184 addRelocationForSymbol(RE, Value.SymbolName); 2185 else 2186 addRelocationForSection(RE, Value.SectionID); 2187 } else if (RelType == ELF::R_X86_64_GOT64) { 2188 // Fill in a 64-bit GOT offset. 2189 uint64_t GOTOffset = allocateGOTEntries(1); 2190 resolveRelocation(Sections[SectionID], Offset, GOTOffset, 2191 ELF::R_X86_64_64, 0); 2192 2193 // Fill in the value of the symbol we're targeting into the GOT 2194 RelocationEntry RE = 2195 computeGOTOffsetRE(GOTOffset, Value.Offset, ELF::R_X86_64_64); 2196 if (Value.SymbolName) 2197 addRelocationForSymbol(RE, Value.SymbolName); 2198 else 2199 addRelocationForSection(RE, Value.SectionID); 2200 } else if (RelType == ELF::R_X86_64_GOTPC32) { 2201 // Materialize the address of the base of the GOT relative to the PC. 2202 // This doesn't create a GOT entry, but it does mean we need a GOT 2203 // section. 2204 (void)allocateGOTEntries(0); 2205 resolveGOTOffsetRelocation(SectionID, Offset, Addend, ELF::R_X86_64_PC32); 2206 } else if (RelType == ELF::R_X86_64_GOTPC64) { 2207 (void)allocateGOTEntries(0); 2208 resolveGOTOffsetRelocation(SectionID, Offset, Addend, ELF::R_X86_64_PC64); 2209 } else if (RelType == ELF::R_X86_64_GOTOFF64) { 2210 // GOTOFF relocations ultimately require a section difference relocation. 2211 (void)allocateGOTEntries(0); 2212 processSimpleRelocation(SectionID, Offset, RelType, Value); 2213 } else if (RelType == ELF::R_X86_64_PC32) { 2214 Value.Addend += support::ulittle32_t::ref(computePlaceholderAddress(SectionID, Offset)); 2215 processSimpleRelocation(SectionID, Offset, RelType, Value); 2216 } else if (RelType == ELF::R_X86_64_PC64) { 2217 Value.Addend += support::ulittle64_t::ref( 2218 computePlaceholderAddress(SectionID, Offset)); 2219 processSimpleRelocation(SectionID, Offset, RelType, Value); 2220 } else if (RelType == ELF::R_X86_64_GOTTPOFF) { 2221 processX86_64GOTTPOFFRelocation(SectionID, Offset, Value, Addend); 2222 } else if (RelType == ELF::R_X86_64_TLSGD || 2223 RelType == ELF::R_X86_64_TLSLD) { 2224 // The next relocation must be the relocation for __tls_get_addr. 2225 ++RelI; 2226 auto &GetAddrRelocation = *RelI; 2227 processX86_64TLSRelocation(SectionID, Offset, RelType, Value, Addend, 2228 GetAddrRelocation); 2229 } else { 2230 processSimpleRelocation(SectionID, Offset, RelType, Value); 2231 } 2232 } else if (Arch == Triple::riscv32 || Arch == Triple::riscv64) { 2233 // *_LO12 relocation receive information about a symbol from the 2234 // corresponding *_HI20 relocation, so we have to collect this information 2235 // before resolving 2236 if (RelType == ELF::R_RISCV_GOT_HI20 || 2237 RelType == ELF::R_RISCV_PCREL_HI20 || 2238 RelType == ELF::R_RISCV_TPREL_HI20 || 2239 RelType == ELF::R_RISCV_TLS_GD_HI20 || 2240 RelType == ELF::R_RISCV_TLS_GOT_HI20) { 2241 RelocationEntry RE(SectionID, Offset, RelType, Addend); 2242 PendingRelocs.push_back({Value, RE}); 2243 } 2244 processSimpleRelocation(SectionID, Offset, RelType, Value); 2245 } else { 2246 if (Arch == Triple::x86) { 2247 Value.Addend += support::ulittle32_t::ref( 2248 computePlaceholderAddress(SectionID, Offset)); 2249 } 2250 processSimpleRelocation(SectionID, Offset, RelType, Value); 2251 } 2252 return ++RelI; 2253 } 2254 2255 void RuntimeDyldELF::processX86_64GOTTPOFFRelocation(unsigned SectionID, 2256 uint64_t Offset, 2257 RelocationValueRef Value, 2258 int64_t Addend) { 2259 // Use the approach from "x86-64 Linker Optimizations" from the TLS spec 2260 // to replace the GOTTPOFF relocation with a TPOFF relocation. The spec 2261 // only mentions one optimization even though there are two different 2262 // code sequences for the Initial Exec TLS Model. We match the code to 2263 // find out which one was used. 2264 2265 // A possible TLS code sequence and its replacement 2266 struct CodeSequence { 2267 // The expected code sequence 2268 ArrayRef<uint8_t> ExpectedCodeSequence; 2269 // The negative offset of the GOTTPOFF relocation to the beginning of 2270 // the sequence 2271 uint64_t TLSSequenceOffset; 2272 // The new code sequence 2273 ArrayRef<uint8_t> NewCodeSequence; 2274 // The offset of the new TPOFF relocation 2275 uint64_t TpoffRelocationOffset; 2276 }; 2277 2278 std::array<CodeSequence, 2> CodeSequences; 2279 2280 // Initial Exec Code Model Sequence 2281 { 2282 static const std::initializer_list<uint8_t> ExpectedCodeSequenceList = { 2283 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 2284 0x00, // mov %fs:0, %rax 2285 0x48, 0x03, 0x05, 0x00, 0x00, 0x00, 0x00 // add x@gotpoff(%rip), 2286 // %rax 2287 }; 2288 CodeSequences[0].ExpectedCodeSequence = 2289 ArrayRef<uint8_t>(ExpectedCodeSequenceList); 2290 CodeSequences[0].TLSSequenceOffset = 12; 2291 2292 static const std::initializer_list<uint8_t> NewCodeSequenceList = { 2293 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00, // mov %fs:0, %rax 2294 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00 // lea x@tpoff(%rax), %rax 2295 }; 2296 CodeSequences[0].NewCodeSequence = ArrayRef<uint8_t>(NewCodeSequenceList); 2297 CodeSequences[0].TpoffRelocationOffset = 12; 2298 } 2299 2300 // Initial Exec Code Model Sequence, II 2301 { 2302 static const std::initializer_list<uint8_t> ExpectedCodeSequenceList = { 2303 0x48, 0x8b, 0x05, 0x00, 0x00, 0x00, 0x00, // mov x@gotpoff(%rip), %rax 2304 0x64, 0x48, 0x8b, 0x00, 0x00, 0x00, 0x00 // mov %fs:(%rax), %rax 2305 }; 2306 CodeSequences[1].ExpectedCodeSequence = 2307 ArrayRef<uint8_t>(ExpectedCodeSequenceList); 2308 CodeSequences[1].TLSSequenceOffset = 3; 2309 2310 static const std::initializer_list<uint8_t> NewCodeSequenceList = { 2311 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00, // 6 byte nop 2312 0x64, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00, // mov %fs:x@tpoff, %rax 2313 }; 2314 CodeSequences[1].NewCodeSequence = ArrayRef<uint8_t>(NewCodeSequenceList); 2315 CodeSequences[1].TpoffRelocationOffset = 10; 2316 } 2317 2318 bool Resolved = false; 2319 auto &Section = Sections[SectionID]; 2320 for (const auto &C : CodeSequences) { 2321 assert(C.ExpectedCodeSequence.size() == C.NewCodeSequence.size() && 2322 "Old and new code sequences must have the same size"); 2323 2324 if (Offset < C.TLSSequenceOffset || 2325 (Offset - C.TLSSequenceOffset + C.NewCodeSequence.size()) > 2326 Section.getSize()) { 2327 // This can't be a matching sequence as it doesn't fit in the current 2328 // section 2329 continue; 2330 } 2331 2332 auto TLSSequenceStartOffset = Offset - C.TLSSequenceOffset; 2333 auto *TLSSequence = Section.getAddressWithOffset(TLSSequenceStartOffset); 2334 if (ArrayRef<uint8_t>(TLSSequence, C.ExpectedCodeSequence.size()) != 2335 C.ExpectedCodeSequence) { 2336 continue; 2337 } 2338 2339 memcpy(TLSSequence, C.NewCodeSequence.data(), C.NewCodeSequence.size()); 2340 2341 // The original GOTTPOFF relocation has an addend as it is PC relative, 2342 // so it needs to be corrected. The TPOFF32 relocation is used as an 2343 // absolute value (which is an offset from %fs:0), so remove the addend 2344 // again. 2345 RelocationEntry RE(SectionID, 2346 TLSSequenceStartOffset + C.TpoffRelocationOffset, 2347 ELF::R_X86_64_TPOFF32, Value.Addend - Addend); 2348 2349 if (Value.SymbolName) 2350 addRelocationForSymbol(RE, Value.SymbolName); 2351 else 2352 addRelocationForSection(RE, Value.SectionID); 2353 2354 Resolved = true; 2355 break; 2356 } 2357 2358 if (!Resolved) { 2359 // The GOTTPOFF relocation was not used in one of the sequences 2360 // described in the spec, so we can't optimize it to a TPOFF 2361 // relocation. 2362 uint64_t GOTOffset = allocateGOTEntries(1); 2363 resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend, 2364 ELF::R_X86_64_PC32); 2365 RelocationEntry RE = 2366 computeGOTOffsetRE(GOTOffset, Value.Offset, ELF::R_X86_64_TPOFF64); 2367 if (Value.SymbolName) 2368 addRelocationForSymbol(RE, Value.SymbolName); 2369 else 2370 addRelocationForSection(RE, Value.SectionID); 2371 } 2372 } 2373 2374 void RuntimeDyldELF::processX86_64TLSRelocation( 2375 unsigned SectionID, uint64_t Offset, uint64_t RelType, 2376 RelocationValueRef Value, int64_t Addend, 2377 const RelocationRef &GetAddrRelocation) { 2378 // Since we are statically linking and have no additional DSOs, we can resolve 2379 // the relocation directly without using __tls_get_addr. 2380 // Use the approach from "x86-64 Linker Optimizations" from the TLS spec 2381 // to replace it with the Local Exec relocation variant. 2382 2383 // Find out whether the code was compiled with the large or small memory 2384 // model. For this we look at the next relocation which is the relocation 2385 // for the __tls_get_addr function. If it's a 32 bit relocation, it's the 2386 // small code model, with a 64 bit relocation it's the large code model. 2387 bool IsSmallCodeModel; 2388 // Is the relocation for the __tls_get_addr a PC-relative GOT relocation? 2389 bool IsGOTPCRel = false; 2390 2391 switch (GetAddrRelocation.getType()) { 2392 case ELF::R_X86_64_GOTPCREL: 2393 case ELF::R_X86_64_REX_GOTPCRELX: 2394 case ELF::R_X86_64_GOTPCRELX: 2395 IsGOTPCRel = true; 2396 [[fallthrough]]; 2397 case ELF::R_X86_64_PLT32: 2398 IsSmallCodeModel = true; 2399 break; 2400 case ELF::R_X86_64_PLTOFF64: 2401 IsSmallCodeModel = false; 2402 break; 2403 default: 2404 report_fatal_error( 2405 "invalid TLS relocations for General/Local Dynamic TLS Model: " 2406 "expected PLT or GOT relocation for __tls_get_addr function"); 2407 } 2408 2409 // The negative offset to the start of the TLS code sequence relative to 2410 // the offset of the TLSGD/TLSLD relocation 2411 uint64_t TLSSequenceOffset; 2412 // The expected start of the code sequence 2413 ArrayRef<uint8_t> ExpectedCodeSequence; 2414 // The new TLS code sequence that will replace the existing code 2415 ArrayRef<uint8_t> NewCodeSequence; 2416 2417 if (RelType == ELF::R_X86_64_TLSGD) { 2418 // The offset of the new TPOFF32 relocation (offset starting from the 2419 // beginning of the whole TLS sequence) 2420 uint64_t TpoffRelocOffset; 2421 2422 if (IsSmallCodeModel) { 2423 if (!IsGOTPCRel) { 2424 static const std::initializer_list<uint8_t> CodeSequence = { 2425 0x66, // data16 (no-op prefix) 2426 0x48, 0x8d, 0x3d, 0x00, 0x00, 2427 0x00, 0x00, // lea <disp32>(%rip), %rdi 2428 0x66, 0x66, // two data16 prefixes 2429 0x48, // rex64 (no-op prefix) 2430 0xe8, 0x00, 0x00, 0x00, 0x00 // call __tls_get_addr@plt 2431 }; 2432 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2433 TLSSequenceOffset = 4; 2434 } else { 2435 // This code sequence is not described in the TLS spec but gcc 2436 // generates it sometimes. 2437 static const std::initializer_list<uint8_t> CodeSequence = { 2438 0x66, // data16 (no-op prefix) 2439 0x48, 0x8d, 0x3d, 0x00, 0x00, 2440 0x00, 0x00, // lea <disp32>(%rip), %rdi 2441 0x66, // data16 prefix (no-op prefix) 2442 0x48, // rex64 (no-op prefix) 2443 0xff, 0x15, 0x00, 0x00, 0x00, 2444 0x00 // call *__tls_get_addr@gotpcrel(%rip) 2445 }; 2446 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2447 TLSSequenceOffset = 4; 2448 } 2449 2450 // The replacement code for the small code model. It's the same for 2451 // both sequences. 2452 static const std::initializer_list<uint8_t> SmallSequence = { 2453 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 2454 0x00, // mov %fs:0, %rax 2455 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00 // lea x@tpoff(%rax), 2456 // %rax 2457 }; 2458 NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); 2459 TpoffRelocOffset = 12; 2460 } else { 2461 static const std::initializer_list<uint8_t> CodeSequence = { 2462 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, 0x00, // lea <disp32>(%rip), 2463 // %rdi 2464 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 2465 0x00, // movabs $__tls_get_addr@pltoff, %rax 2466 0x48, 0x01, 0xd8, // add %rbx, %rax 2467 0xff, 0xd0 // call *%rax 2468 }; 2469 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2470 TLSSequenceOffset = 3; 2471 2472 // The replacement code for the large code model 2473 static const std::initializer_list<uint8_t> LargeSequence = { 2474 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 2475 0x00, // mov %fs:0, %rax 2476 0x48, 0x8d, 0x80, 0x00, 0x00, 0x00, 0x00, // lea x@tpoff(%rax), 2477 // %rax 2478 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00 // nopw 0x0(%rax,%rax,1) 2479 }; 2480 NewCodeSequence = ArrayRef<uint8_t>(LargeSequence); 2481 TpoffRelocOffset = 12; 2482 } 2483 2484 // The TLSGD/TLSLD relocations are PC-relative, so they have an addend. 2485 // The new TPOFF32 relocations is used as an absolute offset from 2486 // %fs:0, so remove the TLSGD/TLSLD addend again. 2487 RelocationEntry RE(SectionID, Offset - TLSSequenceOffset + TpoffRelocOffset, 2488 ELF::R_X86_64_TPOFF32, Value.Addend - Addend); 2489 if (Value.SymbolName) 2490 addRelocationForSymbol(RE, Value.SymbolName); 2491 else 2492 addRelocationForSection(RE, Value.SectionID); 2493 } else if (RelType == ELF::R_X86_64_TLSLD) { 2494 if (IsSmallCodeModel) { 2495 if (!IsGOTPCRel) { 2496 static const std::initializer_list<uint8_t> CodeSequence = { 2497 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, // leaq <disp32>(%rip), %rdi 2498 0x00, 0xe8, 0x00, 0x00, 0x00, 0x00 // call __tls_get_addr@plt 2499 }; 2500 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2501 TLSSequenceOffset = 3; 2502 2503 // The replacement code for the small code model 2504 static const std::initializer_list<uint8_t> SmallSequence = { 2505 0x66, 0x66, 0x66, // three data16 prefixes (no-op) 2506 0x64, 0x48, 0x8b, 0x04, 0x25, 2507 0x00, 0x00, 0x00, 0x00 // mov %fs:0, %rax 2508 }; 2509 NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); 2510 } else { 2511 // This code sequence is not described in the TLS spec but gcc 2512 // generates it sometimes. 2513 static const std::initializer_list<uint8_t> CodeSequence = { 2514 0x48, 0x8d, 0x3d, 0x00, 2515 0x00, 0x00, 0x00, // leaq <disp32>(%rip), %rdi 2516 0xff, 0x15, 0x00, 0x00, 2517 0x00, 0x00 // call 2518 // *__tls_get_addr@gotpcrel(%rip) 2519 }; 2520 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2521 TLSSequenceOffset = 3; 2522 2523 // The replacement is code is just like above but it needs to be 2524 // one byte longer. 2525 static const std::initializer_list<uint8_t> SmallSequence = { 2526 0x0f, 0x1f, 0x40, 0x00, // 4 byte nop 2527 0x64, 0x48, 0x8b, 0x04, 0x25, 2528 0x00, 0x00, 0x00, 0x00 // mov %fs:0, %rax 2529 }; 2530 NewCodeSequence = ArrayRef<uint8_t>(SmallSequence); 2531 } 2532 } else { 2533 // This is the same sequence as for the TLSGD sequence with the large 2534 // memory model above 2535 static const std::initializer_list<uint8_t> CodeSequence = { 2536 0x48, 0x8d, 0x3d, 0x00, 0x00, 0x00, 0x00, // lea <disp32>(%rip), 2537 // %rdi 2538 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 2539 0x48, // movabs $__tls_get_addr@pltoff, %rax 2540 0x01, 0xd8, // add %rbx, %rax 2541 0xff, 0xd0 // call *%rax 2542 }; 2543 ExpectedCodeSequence = ArrayRef<uint8_t>(CodeSequence); 2544 TLSSequenceOffset = 3; 2545 2546 // The replacement code for the large code model 2547 static const std::initializer_list<uint8_t> LargeSequence = { 2548 0x66, 0x66, 0x66, // three data16 prefixes (no-op) 2549 0x66, 0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 2550 0x00, // 10 byte nop 2551 0x64, 0x48, 0x8b, 0x04, 0x25, 0x00, 0x00, 0x00, 0x00 // mov %fs:0,%rax 2552 }; 2553 NewCodeSequence = ArrayRef<uint8_t>(LargeSequence); 2554 } 2555 } else { 2556 llvm_unreachable("both TLS relocations handled above"); 2557 } 2558 2559 assert(ExpectedCodeSequence.size() == NewCodeSequence.size() && 2560 "Old and new code sequences must have the same size"); 2561 2562 auto &Section = Sections[SectionID]; 2563 if (Offset < TLSSequenceOffset || 2564 (Offset - TLSSequenceOffset + NewCodeSequence.size()) > 2565 Section.getSize()) { 2566 report_fatal_error("unexpected end of section in TLS sequence"); 2567 } 2568 2569 auto *TLSSequence = Section.getAddressWithOffset(Offset - TLSSequenceOffset); 2570 if (ArrayRef<uint8_t>(TLSSequence, ExpectedCodeSequence.size()) != 2571 ExpectedCodeSequence) { 2572 report_fatal_error( 2573 "invalid TLS sequence for Global/Local Dynamic TLS Model"); 2574 } 2575 2576 memcpy(TLSSequence, NewCodeSequence.data(), NewCodeSequence.size()); 2577 } 2578 2579 size_t RuntimeDyldELF::getGOTEntrySize() { 2580 // We don't use the GOT in all of these cases, but it's essentially free 2581 // to put them all here. 2582 size_t Result = 0; 2583 switch (Arch) { 2584 case Triple::x86_64: 2585 case Triple::aarch64: 2586 case Triple::aarch64_be: 2587 case Triple::loongarch64: 2588 case Triple::ppc64: 2589 case Triple::ppc64le: 2590 case Triple::systemz: 2591 Result = sizeof(uint64_t); 2592 break; 2593 case Triple::x86: 2594 case Triple::arm: 2595 case Triple::thumb: 2596 Result = sizeof(uint32_t); 2597 break; 2598 case Triple::mips: 2599 case Triple::mipsel: 2600 case Triple::mips64: 2601 case Triple::mips64el: 2602 if (IsMipsO32ABI || IsMipsN32ABI) 2603 Result = sizeof(uint32_t); 2604 else if (IsMipsN64ABI) 2605 Result = sizeof(uint64_t); 2606 else 2607 llvm_unreachable("Mips ABI not handled"); 2608 break; 2609 default: 2610 llvm_unreachable("Unsupported CPU type!"); 2611 } 2612 return Result; 2613 } 2614 2615 uint64_t RuntimeDyldELF::allocateGOTEntries(unsigned no) { 2616 if (GOTSectionID == 0) { 2617 GOTSectionID = Sections.size(); 2618 // Reserve a section id. We'll allocate the section later 2619 // once we know the total size 2620 Sections.push_back(SectionEntry(".got", nullptr, 0, 0, 0)); 2621 } 2622 uint64_t StartOffset = CurrentGOTIndex * getGOTEntrySize(); 2623 CurrentGOTIndex += no; 2624 return StartOffset; 2625 } 2626 2627 uint64_t RuntimeDyldELF::findOrAllocGOTEntry(const RelocationValueRef &Value, 2628 unsigned GOTRelType) { 2629 auto E = GOTOffsetMap.insert({Value, 0}); 2630 if (E.second) { 2631 uint64_t GOTOffset = allocateGOTEntries(1); 2632 2633 // Create relocation for newly created GOT entry 2634 RelocationEntry RE = 2635 computeGOTOffsetRE(GOTOffset, Value.Offset, GOTRelType); 2636 if (Value.SymbolName) 2637 addRelocationForSymbol(RE, Value.SymbolName); 2638 else 2639 addRelocationForSection(RE, Value.SectionID); 2640 2641 E.first->second = GOTOffset; 2642 } 2643 2644 return E.first->second; 2645 } 2646 2647 void RuntimeDyldELF::resolveGOTOffsetRelocation(unsigned SectionID, 2648 uint64_t Offset, 2649 uint64_t GOTOffset, 2650 uint32_t Type) { 2651 // Fill in the relative address of the GOT Entry into the stub 2652 RelocationEntry GOTRE(SectionID, Offset, Type, GOTOffset); 2653 addRelocationForSection(GOTRE, GOTSectionID); 2654 } 2655 2656 RelocationEntry RuntimeDyldELF::computeGOTOffsetRE(uint64_t GOTOffset, 2657 uint64_t SymbolOffset, 2658 uint32_t Type) { 2659 return RelocationEntry(GOTSectionID, GOTOffset, Type, SymbolOffset); 2660 } 2661 2662 void RuntimeDyldELF::processNewSymbol(const SymbolRef &ObjSymbol, SymbolTableEntry& Symbol) { 2663 // This should never return an error as `processNewSymbol` wouldn't have been 2664 // called if getFlags() returned an error before. 2665 auto ObjSymbolFlags = cantFail(ObjSymbol.getFlags()); 2666 2667 if (ObjSymbolFlags & SymbolRef::SF_Indirect) { 2668 if (IFuncStubSectionID == 0) { 2669 // Create a dummy section for the ifunc stubs. It will be actually 2670 // allocated in finalizeLoad() below. 2671 IFuncStubSectionID = Sections.size(); 2672 Sections.push_back( 2673 SectionEntry(".text.__llvm_IFuncStubs", nullptr, 0, 0, 0)); 2674 // First 64B are reserverd for the IFunc resolver 2675 IFuncStubOffset = 64; 2676 } 2677 2678 IFuncStubs.push_back(IFuncStub{IFuncStubOffset, Symbol}); 2679 // Modify the symbol so that it points to the ifunc stub instead of to the 2680 // resolver function. 2681 Symbol = SymbolTableEntry(IFuncStubSectionID, IFuncStubOffset, 2682 Symbol.getFlags()); 2683 IFuncStubOffset += getMaxIFuncStubSize(); 2684 } 2685 } 2686 2687 Error RuntimeDyldELF::finalizeLoad(const ObjectFile &Obj, 2688 ObjSectionToIDMap &SectionMap) { 2689 if (IsMipsO32ABI) 2690 if (!PendingRelocs.empty()) 2691 return make_error<RuntimeDyldError>("Can't find matching LO16 reloc"); 2692 2693 // Create the IFunc stubs if necessary. This must be done before processing 2694 // the GOT entries, as the IFunc stubs may create some. 2695 if (IFuncStubSectionID != 0) { 2696 uint8_t *IFuncStubsAddr = MemMgr.allocateCodeSection( 2697 IFuncStubOffset, 1, IFuncStubSectionID, ".text.__llvm_IFuncStubs"); 2698 if (!IFuncStubsAddr) 2699 return make_error<RuntimeDyldError>( 2700 "Unable to allocate memory for IFunc stubs!"); 2701 Sections[IFuncStubSectionID] = 2702 SectionEntry(".text.__llvm_IFuncStubs", IFuncStubsAddr, IFuncStubOffset, 2703 IFuncStubOffset, 0); 2704 2705 createIFuncResolver(IFuncStubsAddr); 2706 2707 LLVM_DEBUG(dbgs() << "Creating IFunc stubs SectionID: " 2708 << IFuncStubSectionID << " Addr: " 2709 << Sections[IFuncStubSectionID].getAddress() << '\n'); 2710 for (auto &IFuncStub : IFuncStubs) { 2711 auto &Symbol = IFuncStub.OriginalSymbol; 2712 LLVM_DEBUG(dbgs() << "\tSectionID: " << Symbol.getSectionID() 2713 << " Offset: " << format("%p", Symbol.getOffset()) 2714 << " IFuncStubOffset: " 2715 << format("%p\n", IFuncStub.StubOffset)); 2716 createIFuncStub(IFuncStubSectionID, 0, IFuncStub.StubOffset, 2717 Symbol.getSectionID(), Symbol.getOffset()); 2718 } 2719 2720 IFuncStubSectionID = 0; 2721 IFuncStubOffset = 0; 2722 IFuncStubs.clear(); 2723 } 2724 2725 // If necessary, allocate the global offset table 2726 if (GOTSectionID != 0) { 2727 // Allocate memory for the section 2728 size_t TotalSize = CurrentGOTIndex * getGOTEntrySize(); 2729 uint8_t *Addr = MemMgr.allocateDataSection(TotalSize, getGOTEntrySize(), 2730 GOTSectionID, ".got", false); 2731 if (!Addr) 2732 return make_error<RuntimeDyldError>("Unable to allocate memory for GOT!"); 2733 2734 Sections[GOTSectionID] = 2735 SectionEntry(".got", Addr, TotalSize, TotalSize, 0); 2736 2737 // For now, initialize all GOT entries to zero. We'll fill them in as 2738 // needed when GOT-based relocations are applied. 2739 memset(Addr, 0, TotalSize); 2740 if (IsMipsN32ABI || IsMipsN64ABI) { 2741 // To correctly resolve Mips GOT relocations, we need a mapping from 2742 // object's sections to GOTs. 2743 for (section_iterator SI = Obj.section_begin(), SE = Obj.section_end(); 2744 SI != SE; ++SI) { 2745 if (SI->relocation_begin() != SI->relocation_end()) { 2746 Expected<section_iterator> RelSecOrErr = SI->getRelocatedSection(); 2747 if (!RelSecOrErr) 2748 return make_error<RuntimeDyldError>( 2749 toString(RelSecOrErr.takeError())); 2750 2751 section_iterator RelocatedSection = *RelSecOrErr; 2752 ObjSectionToIDMap::iterator i = SectionMap.find(*RelocatedSection); 2753 assert(i != SectionMap.end()); 2754 SectionToGOTMap[i->second] = GOTSectionID; 2755 } 2756 } 2757 GOTSymbolOffsets.clear(); 2758 } 2759 } 2760 2761 // Look for and record the EH frame section. 2762 ObjSectionToIDMap::iterator i, e; 2763 for (i = SectionMap.begin(), e = SectionMap.end(); i != e; ++i) { 2764 const SectionRef &Section = i->first; 2765 2766 StringRef Name; 2767 Expected<StringRef> NameOrErr = Section.getName(); 2768 if (NameOrErr) 2769 Name = *NameOrErr; 2770 else 2771 consumeError(NameOrErr.takeError()); 2772 2773 if (Name == ".eh_frame") { 2774 UnregisteredEHFrameSections.push_back(i->second); 2775 break; 2776 } 2777 } 2778 2779 GOTOffsetMap.clear(); 2780 GOTSectionID = 0; 2781 CurrentGOTIndex = 0; 2782 2783 return Error::success(); 2784 } 2785 2786 bool RuntimeDyldELF::isCompatibleFile(const object::ObjectFile &Obj) const { 2787 return Obj.isELF(); 2788 } 2789 2790 void RuntimeDyldELF::createIFuncResolver(uint8_t *Addr) const { 2791 if (Arch == Triple::x86_64) { 2792 // The adddres of the GOT1 entry is in %r11, the GOT2 entry is in %r11+8 2793 // (see createIFuncStub() for details) 2794 // The following code first saves all registers that contain the original 2795 // function arguments as those registers are not saved by the resolver 2796 // function. %r11 is saved as well so that the GOT2 entry can be updated 2797 // afterwards. Then it calls the actual IFunc resolver function whose 2798 // address is stored in GOT2. After the resolver function returns, all 2799 // saved registers are restored and the return value is written to GOT1. 2800 // Finally, jump to the now resolved function. 2801 // clang-format off 2802 const uint8_t StubCode[] = { 2803 0x57, // push %rdi 2804 0x56, // push %rsi 2805 0x52, // push %rdx 2806 0x51, // push %rcx 2807 0x41, 0x50, // push %r8 2808 0x41, 0x51, // push %r9 2809 0x41, 0x53, // push %r11 2810 0x41, 0xff, 0x53, 0x08, // call *0x8(%r11) 2811 0x41, 0x5b, // pop %r11 2812 0x41, 0x59, // pop %r9 2813 0x41, 0x58, // pop %r8 2814 0x59, // pop %rcx 2815 0x5a, // pop %rdx 2816 0x5e, // pop %rsi 2817 0x5f, // pop %rdi 2818 0x49, 0x89, 0x03, // mov %rax,(%r11) 2819 0xff, 0xe0 // jmp *%rax 2820 }; 2821 // clang-format on 2822 static_assert(sizeof(StubCode) <= 64, 2823 "maximum size of the IFunc resolver is 64B"); 2824 memcpy(Addr, StubCode, sizeof(StubCode)); 2825 } else { 2826 report_fatal_error( 2827 "IFunc resolver is not supported for target architecture"); 2828 } 2829 } 2830 2831 void RuntimeDyldELF::createIFuncStub(unsigned IFuncStubSectionID, 2832 uint64_t IFuncResolverOffset, 2833 uint64_t IFuncStubOffset, 2834 unsigned IFuncSectionID, 2835 uint64_t IFuncOffset) { 2836 auto &IFuncStubSection = Sections[IFuncStubSectionID]; 2837 auto *Addr = IFuncStubSection.getAddressWithOffset(IFuncStubOffset); 2838 2839 if (Arch == Triple::x86_64) { 2840 // The first instruction loads a PC-relative address into %r11 which is a 2841 // GOT entry for this stub. This initially contains the address to the 2842 // IFunc resolver. We can use %r11 here as it's caller saved but not used 2843 // to pass any arguments. In fact, x86_64 ABI even suggests using %r11 for 2844 // code in the PLT. The IFunc resolver will use %r11 to update the GOT 2845 // entry. 2846 // 2847 // The next instruction just jumps to the address contained in the GOT 2848 // entry. As mentioned above, we do this two-step jump by first setting 2849 // %r11 so that the IFunc resolver has access to it. 2850 // 2851 // The IFunc resolver of course also needs to know the actual address of 2852 // the actual IFunc resolver function. This will be stored in a GOT entry 2853 // right next to the first one for this stub. So, the IFunc resolver will 2854 // be able to call it with %r11+8. 2855 // 2856 // In total, two adjacent GOT entries (+relocation) and one additional 2857 // relocation are required: 2858 // GOT1: Address of the IFunc resolver. 2859 // GOT2: Address of the IFunc resolver function. 2860 // IFuncStubOffset+3: 32-bit PC-relative address of GOT1. 2861 uint64_t GOT1 = allocateGOTEntries(2); 2862 uint64_t GOT2 = GOT1 + getGOTEntrySize(); 2863 2864 RelocationEntry RE1(GOTSectionID, GOT1, ELF::R_X86_64_64, 2865 IFuncResolverOffset, {}); 2866 addRelocationForSection(RE1, IFuncStubSectionID); 2867 RelocationEntry RE2(GOTSectionID, GOT2, ELF::R_X86_64_64, IFuncOffset, {}); 2868 addRelocationForSection(RE2, IFuncSectionID); 2869 2870 const uint8_t StubCode[] = { 2871 0x4c, 0x8d, 0x1d, 0x00, 0x00, 0x00, 0x00, // leaq 0x0(%rip),%r11 2872 0x41, 0xff, 0x23 // jmpq *(%r11) 2873 }; 2874 assert(sizeof(StubCode) <= getMaxIFuncStubSize() && 2875 "IFunc stub size must not exceed getMaxIFuncStubSize()"); 2876 memcpy(Addr, StubCode, sizeof(StubCode)); 2877 2878 // The PC-relative value starts 4 bytes from the end of the leaq 2879 // instruction, so the addend is -4. 2880 resolveGOTOffsetRelocation(IFuncStubSectionID, IFuncStubOffset + 3, 2881 GOT1 - 4, ELF::R_X86_64_PC32); 2882 } else { 2883 report_fatal_error("IFunc stub is not supported for target architecture"); 2884 } 2885 } 2886 2887 unsigned RuntimeDyldELF::getMaxIFuncStubSize() const { 2888 if (Arch == Triple::x86_64) { 2889 return 10; 2890 } 2891 return 0; 2892 } 2893 2894 bool RuntimeDyldELF::relocationNeedsGot(const RelocationRef &R) const { 2895 unsigned RelTy = R.getType(); 2896 if (Arch == Triple::aarch64 || Arch == Triple::aarch64_be) 2897 return RelTy == ELF::R_AARCH64_ADR_GOT_PAGE || 2898 RelTy == ELF::R_AARCH64_LD64_GOT_LO12_NC; 2899 2900 if (Arch == Triple::loongarch64) 2901 return RelTy == ELF::R_LARCH_GOT_PC_HI20 || 2902 RelTy == ELF::R_LARCH_GOT_PC_LO12; 2903 2904 if (Arch == Triple::x86_64) 2905 return RelTy == ELF::R_X86_64_GOTPCREL || 2906 RelTy == ELF::R_X86_64_GOTPCRELX || 2907 RelTy == ELF::R_X86_64_GOT64 || 2908 RelTy == ELF::R_X86_64_REX_GOTPCRELX; 2909 return false; 2910 } 2911 2912 bool RuntimeDyldELF::relocationNeedsStub(const RelocationRef &R) const { 2913 if (Arch != Triple::x86_64) 2914 return true; // Conservative answer 2915 2916 switch (R.getType()) { 2917 default: 2918 return true; // Conservative answer 2919 2920 2921 case ELF::R_X86_64_GOTPCREL: 2922 case ELF::R_X86_64_GOTPCRELX: 2923 case ELF::R_X86_64_REX_GOTPCRELX: 2924 case ELF::R_X86_64_GOTPC64: 2925 case ELF::R_X86_64_GOT64: 2926 case ELF::R_X86_64_GOTOFF64: 2927 case ELF::R_X86_64_PC32: 2928 case ELF::R_X86_64_PC64: 2929 case ELF::R_X86_64_64: 2930 // We know that these reloation types won't need a stub function. This list 2931 // can be extended as needed. 2932 return false; 2933 } 2934 } 2935 2936 } // namespace llvm 2937